Paper deep dive
Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning
Rongzhe Wei, Mufei Li, Mohsen Ghassemi, Eleonora Kreačić, Yifan Li, Xiang Yue, Bo Li, Vamsi K. Potluru, Pan Li, Eli Chien
Models: Large language models (multiple scales, unspecified)
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 3/12/2026, 6:33:25 PM
Summary
The paper identifies a critical blind spot in current Large Language Model (LLM) unlearning evaluation frameworks, which typically rely on random data removal. The authors demonstrate that minority populations, whose data are often treated as outliers, exhibit higher memorization propensity and are significantly harder to unlearn. By introducing a minority-aware evaluation framework using PII-based canaries, the study reveals that these groups suffer at least 20% higher privacy leakage compared to average-case assessments, advocating for more equitable and robust privacy auditing.
Entities (6)
Relation Signals (3)
Machine Unlearning → appliedto → Large Language Models
confidence 100% · Machine unlearning seeks to efficiently modify the LLM so that it becomes statistically indistinguishable from a model retrained from scratch.
Membership Inference Attacks → evaluates → Machine Unlearning
confidence 98% · MIAs ... have been widely adopted for assessing unlearning performance.
Minority-aware evaluation framework → identifiesriskfor → Minority populations
confidence 95% · The proposed minority-aware evaluation framework marks an essential step toward more equitable and comprehensive assessments of LLM unlearning efficacy.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Language Models (LLMs) embed sensitive, human-generated data, prompting the need for unlearning methods. Although certified unlearning offers strong privacy guarantees, its restrictive assumptions make it unsuitable for LLMs, giving rise to various heuristic approaches typically assessed through empirical evaluations. These standard evaluations randomly select data for removal, apply unlearning techniques, and use membership inference attacks (MIAs) to compare unlearned models against models retrained without the removed data. However, to ensure robust privacy protections for every data point, it is essential to account for scenarios in which certain data subsets face elevated risks. Prior research suggests that outliers, particularly including data tied to minority groups, often exhibit higher memorization propensity which indicates they may be more difficult to unlearn. Building on these insights, we introduce a complementary, minority-aware evaluation framework to highlight blind spots in existing frameworks. We substantiate our findings with carefully designed experiments, using canaries with personally identifiable information (PII) to represent these minority subsets and demonstrate that they suffer at least 20% higher privacy leakage across various unlearning methods, MIAs, datasets, and LLM scales. Our proposed minority-aware evaluation framework marks an essential step toward more equitable and comprehensive assessments of LLM unlearning efficacy.
Tags
Links
- Source: https://arxiv.org/abs/2412.08559
- Canonical: https://arxiv.org/abs/2412.08559
Trouble viewing inline? Open PDF directly →
Full Text
113,014 characters extracted from source content.
Expand or collapse full text
Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning Rongzhe Wei Mufei Li Mohsen Ghassemi Eleonora Kreačić Yifan Li Xiang Yue Bo Li Vamsi K. Potluru Pan Li Eli Chien Abstract Large Language Models (LLMs) embed sensitive, human-generated data, prompting the need for unlearning methods. Although certified unlearning offers strong privacy guarantees, its restrictive assumptions make it unsuitable for LLMs, giving rise to various heuristic approaches typically assessed through empirical evaluations. These standard evaluations randomly select data for removal, apply unlearning techniques, and use membership inference attacks (MIAs) to compare unlearned models against models retrained without the removed data. However, to ensure robust privacy protections for every data point, it is essential to account for scenarios in which certain data subsets face elevated risks. Prior research suggests that outliers, particularly including data tied to minority groups, often exhibit higher memorization propensity which indicates they may be more difficult to unlearn. Building on these insights, we introduce a complementary, minority-aware evaluation framework to highlight blind spots in existing frameworks. We substantiate our findings with carefully designed experiments, using canaries with personally identifiable information (PII) to represent these minority subsets and demonstrate that they suffer at least 20% higher privacy leakage across various unlearning methods, MIAs, datasets, and LLM scales. Our proposed minority-aware evaluation framework111Code source: https://github.com/Graph-COM/Minority_Aware_LLM_Unlearning.git marks an essential step toward more equitable and comprehensive assessments of LLM unlearning efficacy. Machine Learning, ICML .tocmtchapter 1 Introduction Large Language Models (LLMs) are trained on vast and diverse datasets, often sourced from public content on the web, much of which is generated by humans (Touvron et al., 2023; Ouyang et al., 2022). This practice raises significant ethical concerns, particularly when the data includes sensitive information, leading to potential privacy violations. Individuals whose data has been used may seek to exercise their “right to be forgotten”, a protection guaranteed by regulations such as the General Data Protection Regulation (GDPR) (Krzysztofek, 2018). The ideal approach to fulfilling such a request is to retrain the LLM from scratch, excluding the data to be removed. However, this solution is prohibitively expensive and impractical for large-scale models. To address this, the concept of machine unlearning has emerged as a promising alternative. Machine unlearning seeks to efficiently modify the LLM so that it becomes statistically indistinguishable from a model retrained from scratch. In this way, no adversary could confidently determine whether a model has undergone an unlearning process or been retrained, ensuring compliance with the “right to be forgotten”. Unfortunately, it remains an open problem to enforce the formal unlearning guarantee for deep neural networks and LLMs without exact retraining. Despite recent progress in theoretical unlearning research (Guo et al., 2020; Sekhari et al., 2021; Neel et al., 2021; Ullah et al., 2021; Chien et al., 2023; Ullah & Arora, 2023; Chien et al., 2024a, b), their restrictive assumptions limit practical applicability to deep neural networks and LLMs. Concurrently, researchers have developed efficient unlearning heuristics and empirically evaluated their efficacy (Golatkar et al., 2020a, b; Graves et al., 2021; Liu et al., 2024a, c; Yao et al., 2024), often by comparing approximately unlearned models to those retrained from scratch (Pawelczyk et al., 2024a). Among the various evaluation methods, membership inference attacks (MIAs) (Shokri et al., 2017), originally developed to infer data usage during training, have been widely adopted for assessing unlearning performance (Shi et al., 2024b). We identify a critical pitfall in the aforementioned LLM unlearning efficacy evaluation. Prior work has demonstrated that unlearning difficulty can vary substantially across individual data points in vision tasks (Thudi et al., 2024; Zhao et al., 2024). Similarly, recent studies on LLMs also reveal highly non-uniform memorization patterns, where certain samples are memorized much more strongly than others (Feldman & Zhang, 2020; Carlini et al., 2022). However, current unlearning evaluation methods only capture “average-case” performance through random data removal from the training set. This approach inadequately addresses privacy risks for hard-to-unlearn data, failing to account for challenging scenarios necessitating rigorous privacy protection (Steinke & Ullman, 2020; Aerni et al., 2024). It neglects the principle that every individual’s right to be forgotten should be upheld equally, thus ignoring data from minority groups, which are often treated as outliers and can be more resistant to unlearning due to the aforementioned stronger memorization effects (Carlini et al., 2022; Nasr et al., 2021, 2023). Consequently, standard unlearning evaluation significantly underestimates privacy risks for these groups, overlooking crucial social responsibilities in personal data protection. Contributions. Motivated by the privacy auditing literature (Jagielski et al., 2020; Steinke et al., 2024), we conduct a synthetic experiment on unlearning injected canaries pertaining to minority groups. We choose Personally Identifiable Information (PII) as a representative minority identifier, while noting that our approach extends to broader cases. We show that minorities suffer from at least 20%percent2020\%20 % more privacy leakage in most cases across combinations of six unlearning approaches, three MIA variants, three datasets, and two LLMs of different scales. These results underscore the prevalence of the issue in practical settings, highlighting the need for a more effective LLM unlearning evaluation, particularly in regard to privacy risks for minority groups. Accordingly, we propose a minority-aware LLM unlearning evaluation protocol (Figure 1) as an initial step toward this goal. With this minority-aware protocol, we benchmark existing unlearning approaches and investigate the effects of forget set size as well as unlearning complexity. This study provides a more holistic understanding of different LLM unlearning approaches for practitioners. Notably, we observe that Langevin Unlearning—the only approach incorporating noise—achieves a favorable privacy-utility trade-off compared to noiseless methods such as SCRUB and Gradient Ascent (GA), suggesting a potential crucial role of noise incorporation in effective unlearning. In summary, these insights underline the critical role of our minority-aware evaluation framework in advancing equitable assessments of unlearning efficacy across different methods. Figure 1: Illustration of our proposed LLM unlearning approaches (highlighted in red), when compared with the existing pipeline. Standard LLM unlearning evaluation typically involves randomly sampling data for removal from the training set (Case 1), which may underestimate privacy leakage for minority groups. In contrast, we design experiments to assess unlearning efficacy by removing canaries (deliberately inserted data points) related to minority groups (Case 2) and by directly removing data from minority groups (Case 3). Our approach provides a more comprehensive, minority-aware evaluation by considering the worst result across the three settings. 2 Related Work Privacy auditing is a fundamental yet challenging aspect of LLM unlearning due to the difficulty of distinguishing training samples effectively (Duan et al., 2024). Various privacy-related metrics, such as exposure (Carlini et al., 2019), mean reciprocal rank (Wu et al., 2023), extraction likelihood (Jang et al., 2022), and truth ratio (Maini et al., 2024), have been proposed to probe privacy leakage. Among these, MIAs remain one of the most crucial tools for evaluating machine unlearning methods (Liu et al., 2024b). Standard MIAs typically involve training numerous shadow models independently to empirically approximate the distribution (Carlini et al., 2022). This approach has also been adopted for LLM unlearning, as seen in the NeurIPS 2023 Machine Unlearning Challenge222https://unlearning-challenge.github.io/assets/data/Machine_Unlearning_Metric.pdf (which compares the point-wise output distributions of multiple unlearned and retrained models to perform MIAs) and Kurmanji et al. (2024); Pawelczyk et al. (2024b). Hayes et al. (2024) further highlights the limitations of average-case evaluations and introduces a specialized per-sample MIA method for unlearning evaluation. Their approach focuses on unlearning a randomly selected subset of training data by training a series of shadow models and performing per-sample MIA using a likelihood ratio test under Gaussian fitting (Carlini et al., 2022). A major downside of MIA approaches involving shadow models is their computational expense, as they require training a large number of LLMs independently (Liu et al., 2024b). To address this downside, another line of research compares the outputs of models using different statistical metrics without requiring shadow models (Zhang et al., 2024; Liu et al., 2024a, c; Yao et al., 2024; Li et al., 2024a), making these methods more computationally feasible (Maini et al., 2024). For instance, Shi et al. (2024b) measures privacy risk through the normalized AUC difference between unlearned and retrained models, using MIAs such as Min-K% (Shi et al., 2024a). However, these works typically select the forget set randomly from the training set, corresponding to an average-case evaluation. Our study highlights a critical limitation of this approach: the privacy risks of minority populations within the training set are severely underestimated because minority data are less likely to be selected in the unlearning evaluation pipeline. By focusing on minority-aware scenarios, our work provides a more comprehensive perspective on unlearning evaluation and privacy risks. In a related context, Zhao et al. (2024) explores how the memorization of image representations affects the difficulty of unlearning individual samples in vision tasks by introducing an entanglement-based metric. Our work further explores the systematic degradation in unlearning efficacy for minority subgroups during the fine-tuning of large language models. As a future direction, it would be valuable to examine the relationship between the entanglement-based difficulty metrics proposed by Zhao et al. (2024) and the minority-aware phenomena we identify in natural language contexts. 3 Preliminaries Machine unlearning (Cao & Yang, 2015; Bourtoule et al., 2021) has emerged as an important direction in trustworthy language models. It was initially motivated by privacy due to “the right to be forgotten” from GDPR and later extended to other legal and ethical concerns, including copyright (Yao et al., 2024), biased or outdated information mitigation (Liu et al., 2024b), hallucination removal (Yao et al., 2023), entity forgetting (Maini et al., 2024) and data poisoning removal (Pawelczyk et al., 2024a; Li et al., 2024b). In this work, we focus on the privacy aspect of the problem, albeit our methodology extends to other cases whenever the indistinguishability to the retrained model is an appropriate metric. We briefly state the generic machine unlearning setting for privacy. Assume a training dataset DtrainsubscripttrainD_trainDtrain and a holdout test set DtestsubscripttestD_testDtest are given. Let Mlearn←(M0,Dtrain)←subscriptlearnsubscript0subscripttrainM_learn (M_0,D_train)Mlearn ← A ( M0 , Dtrain ) be the language model trained on DtrainsubscripttrainD_trainDtrain starting from an initial model M0subscript0M_0M0 via the training algorithm AA, which may be either a pre-trained language model or random initialization. Once the model is trained, we receive data removal requests that partition the training set Dtrain=Dforget∪DkeepsubscripttrainsubscriptforgetsubscriptkeepD_train=D_forget∪ D_keepDtrain = Dforget ∪ Dkeep into a subset to be forgotten later DforgetsubscriptforgetD_forgetDforget and a keep set DkeepsubscriptkeepD_keepDkeep. An unlearning algorithm UU takes MlearnsubscriptlearnM_learnMlearn, DforgetsubscriptforgetD_forgetDforget and DtrainsubscripttrainD_trainDtrain as input to return an updated model Munlearn←(Mlearn,Dforget,Dtrain)←subscriptunlearnsubscriptlearnsubscriptforgetsubscripttrainM_unlearn (M_learn,D_forget,D_% train)Munlearn ← U ( Mlearn , Dforget , Dtrain ). It is worth noting that MunlearnsubscriptunlearnM_unlearnMunlearn depends on the choice of DforgetsubscriptforgetD_forgetDforget. The gold standard to adhere to “the right to be forgotten” is retraining without DforgetsubscriptforgetD_forgetDforget, namely Mretrain←(M0,Dtrain∖Dforget)←subscriptretrainsubscript0subscripttrainsubscriptforgetM_retrain (M_0,D_train D_% forget)Mretrain ← A ( M0 , Dtrain ∖ Dforget ). We say UU achieves good unlearning efficacy if MunlearnsubscriptunlearnM_unlearnMunlearn and MretrainsubscriptretrainM_retrainMretrain are indistinguishable in their behavior m(Munlearn,D)≈m(Mretrain,D)subscriptunlearnsubscriptretrainm(M_unlearn,D)≈ m(M_retrain,D)m ( Munlearn , D ) ≈ m ( Mretrain , D ) on any corpus D, where m is any evaluation metric. Since MunlearnsubscriptunlearnM_unlearnMunlearn and MretrainsubscriptretrainM_retrainMretrain depend on the choice DforgetsubscriptforgetD_forgetDforget, such approximation should be taken over the worst case ideally. 3.1 Efficient MIAs for LLM Unlearning As previously discussed, the effectiveness of unlearning methods can be measured by the indistinguishability between the resulting unlearned models and an exactly retrained model. MIA is often leveraged to determine whether a specific sample is part of the training set and is widely applied to audit training data privacy leakage. Therefore, to evaluate the efficacy of an unlearning approach, we consider the PrivLeak (PL) metric (Shi et al., 2024b). First, we define the difference in AUC scores between the unlearned model MunlearnsubscriptunlearnM_unlearnMunlearn and the retrained model MretrainsubscriptretrainM_retrainMretrain as ΔAUC=AUC(Munlearn;Dforget,Dtest)−AUC(Mretrain;Dforget,Dtest)ΔAUCAUCsubscriptunlearnsubscriptforgetsubscripttestAUCsubscriptretrainsubscriptforgetsubscripttest =AUC(M_unlearn;D_forget,D_test% )-AUC(M_retrain;D_forget,D_test)Δ AUC = AUC ( Munlearn ; Dforget , Dtest ) - AUC ( Mretrain ; Dforget , Dtest ), where AUC is the AUC-ROC score of an MIA (Ye et al., 2022) that tries to discriminate samples from DforgetsubscriptforgetD_forgetDforget and DtestsubscripttestD_testDtest based on the output statistics (e.g. loss) of a given model M. Then, the PL metric is formulated as: PrivLeak (PL)=ΔAUCAUC(Mretrain;Dforget,Dtest)PrivLeak (PL)ΔAUCAUCsubscriptretrainsubscriptforgetsubscripttest -2mPrivLeak (PL)= % AUC (M_retrain;D_forget,D_test )% -2mPrivLeak (PL) = divide start_ARG Δ AUC end_ARG start_ARG AUC ( Mretrain ; Dforget , Dtest ) end_ARG (1) By normalizing the difference in AUC scores between MunlearnsubscriptunlearnM_unlearnMunlearn and MretrainsubscriptretrainM_retrainMretrain using the AUC of MretrainsubscriptretrainM_retrainMretrain, the metric accounts for the inherent difficulty of distinguishing the forget and test sets. Note that for an effective unlearning method, the metric should be around zero since the behavior of Munlearn,MretrainsubscriptunlearnsubscriptretrainM_unlearn,M_retrainMunlearn , Mretrain are indistinguishable. A larger magnitude of the PL metric implies a greater amount of privacy information that has been leaked under the tested MIA. A positive value indicates that the sample has not been fully forgotten, as the attacker has a higher AUC for MunlearnsubscriptunlearnM_unlearnMunlearn than MretrainsubscriptretrainM_retrainMretrain. Conversely, a negative metric value suggests over-forgetting, which still indicates that MunlearnsubscriptunlearnM_unlearnMunlearn differs from MretrainsubscriptretrainM_retrainMretrain and thus cause privacy breaches. Finally, note that an effective unlearning solution should lead to a small PL metric for any choice of MIA. In this work, we consider three popular MIAs and report the corresponding PL metric. • lossMIA (Yeom et al., 2018): Determines membership of a sample x for a model M based on its loss ℓ(M;x)ℓ (M;x)ℓ ( M ; x ). • zlibMIA (Carlini et al., 2021): Determines membership of a sample based on the sample loss normalized by its zlib compression size, ℓ(M;x)/zlib(x)ℓzlib (M;x)/zlib(x)ℓ ( M ; x ) / zlib ( x ). • Min-KK% (Shi et al., 2023): Selects the lowest K%percentK\%K % of token likelihoods and leverages the corresponding negative log-likelihood for membership inference. 4 The Underestimated Privacy Risk of Data Minorities Recall that both the unlearned Munlearn←(Mlearn,Dforget,Dtrain)←subscriptunlearnsubscriptlearnsubscriptforgetsubscripttrainM_unlearn (M_learn,D_forget,D_% train)Munlearn ← U ( Mlearn , Dforget , Dtrain ) and retrained Mretrain←(M0,Dtrain∖Dforget)←subscriptretrainsubscript0subscripttrainsubscriptforgetM_retrain (M_0,D_train D_% forget)Mretrain ← A ( M0 , Dtrain ∖ Dforget ) language models depend on the choice of the forget set DforgetsubscriptforgetD_forgetDforget. Whenever we estimate the privacy leakage of an unlearning method UU via some evaluation m, it is important to account for potential high-risk partitions of DforgetsubscriptforgetD_forgetDforget to ensure a comprehensive assessment of privacy risk. Unfortunately, the current LLM unlearning evaluation pipeline overlooks this critical aspect, where the partition leading to DforgetsubscriptforgetD_forgetDforget is chosen uniformly at random (Jang et al., 2023; Chen & Yang, 2023; Yao et al., 2024; Maini et al., 2024; Zhang et al., 2024; Shi et al., 2024b). The reported privacy risk therein hence corresponds to the “average case”, which may significantly underestimate the privacy risk of highly privacy-sensitive points that request unlearning. It is known in the privacy literature that some rare training samples (minorities) may have an outsized effect on model memorization compared to common training samples (majorities) (Feldman & Zhang, 2020; Carlini et al., 2022). Intuitively, a similar phenomenon persists for unlearning. Table 1: Top three most frequent and least frequent area codes within Enron dataset. Area code Count 713 (Houston) 135,307 800 (Toll-free) 11,902 212 (New York) 10,739 484 (Allentown) 1 Here we utilize the Enron dataset as a case study. This dataset comprises 535,703 authentic emails from 158 employees of the Enron Corporation. It is a standard benchmark dataset for studying PII leakage, where the phone number is one form of PII that has been extensively studied (Lukas et al., 2023). The phone numbers here follow the format of the U.S. phone numbers (e.g., 123-456-7890), with the first three digits serving as the area code, representing the location where the number holder applied for the number. Such information is considered sensitive as it leaks not only the phone number itself, but also the geographic information pertaining to the number holder. Table 1 illustrates the least frequent and three most frequent area codes in the Enron dataset. The area code distribution is far from uniform. Consequently, if emails containing phone numbers are uniformly sampled for the forget set DforgetsubscriptforgetD_forgetDforget, Figure 2: Area code histogram in Enron Dataset. minority data, such as emails with rare area codes like 484, are unlikely to be included due to their lower frequency. If unlearning minority data is inherently more challenging and results in greater privacy leakage, the existing evaluation pipeline may underestimate privacy risks for minorities. 4.1 Verify Underestimated Privacy Risks of Minority via Canary Injection Figure 3: Illustration of the forget set DforgetsubscriptforgetD_forgetDforget, the construction of the canary set DcanarysubscriptcanaryD_canaryDcanary, and the minority set DminoritysubscriptminorityD_minorityDminority for the Enron dataset. The minority set consists of emails with phone numbers containing the least frequent area codes. A histogram and distribution of area codes (e.g., 713 as the most frequent and 484 as the least frequent) are shown in Figure 2 and Table 1. To rigorously show that removing data from minority333Minority here refers to any subset of the data, defined by some shared value, that is under-represented in the training set. This term is generic and may apply to any type of attribute, demographic or otherwise. populations indeed leads to higher unlearning privacy leakage, we design experiments based on the idea of canaries in the privacy auditing literature (Jagielski et al., 2020; Steinke et al., 2024). For simplicity, we focus on the scenario where data removal requests pertain to PIIs, where each training sample x∈Dtrain(1)superscriptsubscripttrain1x∈ D_train^(1)x ∈ Dtrain( 1 ) consists of PIIs such as phone numbers or organizations. We choose PIIs as a representative minority identifier, albeit a similar idea extends beyond PIIs. We consider the following cases, see Figure 1 for an illustration. 1) Random: we randomly partition Dtrain(1)=Dforget∪Dkeepsuperscriptsubscripttrain1subscriptforgetsubscriptkeepD_train^(1)=D_forget∪ D_keepDtrain( 1 ) = Dforget ∪ Dkeep as in the standard unlearning evaluation pipeline. This leads to Mlearn(1)←(M0,Dtrain(1))←superscriptsubscriptlearn1subscript0superscriptsubscripttrain1M_learn^(1) (M_0,D_train^(1))Mlearn( 1 ) ← A ( M0 , Dtrain( 1 ) ). 2) Canary: For the same forget set Dforget=xii=1nsubscriptforgetsuperscriptsubscriptsubscript1D_forget=\x_i\_i=1^nDforget = xitalic_i i = 1n, we construct a canary set Dcanary=xi′i=1nsubscriptcanarysuperscriptsubscriptsuperscriptsubscript′1D_canary=\x_i \_i=1^nDcanary = xitalic_i′ i = 1n, where each xi′subscript′x_i xitalic_i′ is identical to xisubscriptx_ixitalic_i except that only the PII is replaced by the least frequent one among Dtrain(1)superscriptsubscripttrain1D_train^(1)Dtrain( 1 ), isolating the impact of non-PII components. Finally, we construct a synthetic training set Dtrain(2)=Dcanary∪Dkeepsuperscriptsubscripttrain2subscriptcanarysubscriptkeepD_train^(2)=D_canary∪ D_keepDtrain( 2 ) = Dcanary ∪ Dkeep, which leads to Mlearn(2)←(M0,Dtrain(2))←subscriptsuperscript2learnsubscript0superscriptsubscripttrain2M^(2)_learn (M_0,D_train^(2))M( 2 )learn ← A ( M0 , Dtrain( 2 ) ). By executing the same unlearning evaluation process for both cases Mlearn(1)subscriptsuperscript1learnM^(1)_learnM( 1 )learn and Mlearn(2)subscriptsuperscript2learnM^(2)_learnM( 2 )learn, we aim to show that the privacy risk for Canary is much higher than Random. By applying the same unlearning algorithm for removing DforgetsubscriptforgetD_forgetDforget and DcanarysubscriptcanaryD_canaryDcanary, we obtain the unlearned model Munlearn(1)superscriptsubscriptunlearn1M_unlearn^(1)Munlearn( 1 ) and Munlearn(2)superscriptsubscriptunlearn2M_unlearn^(2)Munlearn( 2 ) respectively. The privacy leakage (PL) is then computed for these cases as described in Section 3.1. The calculation of PL for Canary entails replacing DforgetsubscriptforgetD_forgetDforget with DcanarysubscriptcanaryD_canaryDcanary in Eq. (1). Note that the retrained model Mretrain←(M0,Dkeep)←subscriptretrainsubscript0subscriptkeepM_retrain (M_0,D_keep)Mretrain ← A ( M0 , Dkeep ) is identical for both scenarios. An illustrative example of canary construction is provided in Figure 3. Note that for each email within DforgetsubscriptforgetD_forgetDforget in Random, we construct the corresponding canary by only replacing its area code with the least frequent one (i.e., 484). This design is critical as we ensure the other part of the email is identical to the original email. Hence, if the privacy leakage of Canary is greater than Random, it must be due to the difference in the area code. We repeat the similar canary construction for the other PII such as email domain and year of legal judgment for different datasets. 4.2 Quantify the Underestimated Privacy Risk of Unlearning Minority While our synthetic experiment on canary injection may be used to verify whether the unlearning privacy risk of minority populations is underestimated in the standard LLM unlearning evaluation pipeline, it cannot quantify the privacy risk for minorities in the real-world setting. We further design the third case aiming at quantifying the amount of underestimated privacy risk by directly choosing data to be removed containing the least frequent PII. 3) Minority: construct a set DminoritysubscriptminorityD_minorityDminority that is of the same size as DforgetsubscriptforgetD_forgetDforget in Random, which consists of samples with the least frequent PII within the dataset. By comparing the computed privacy risk of Random and Minority, we can quantify the amount of underestimated privacy risk for data removal from minority groups compared to the average case. If the resulting privacy risk is significantly higher than Random, any conclusion pertaining to unlearning efficacy drawn from Random can be misleading and the right to be forgotten of minorities is overlooked. 5 Unlearning Methods We evaluate the following popular unlearning approaches in the literature. With a slight abuse of notation, we denote M for both the model and its parameters for simplicity. • Random Labels (RL) (Golatkar et al., 2020a; Yao et al., 2024): In the context of next-token prediction, this method involves randomly selecting tokens from the entire vocabulary during training on DforgetsubscriptforgetD_forgetDforget, aiming to disturb the model’s learning from this dataset. The intuition behind this approach is that a model uninformed by DforgetsubscriptforgetD_forgetDforget should behave as though it is randomly guessing the next token. However, as argued in Yao et al. (2024), this intuition may not always hold, depending on the specific scenario. • Exact Unlearning (EUk) and Catastrophic Forgetting (CFk) (Goel et al., 2022): Exact unlearning can be done by retraining the entire model from scratch on DkeepsubscriptkeepD_keepDkeep, albeit it is prohibitively expansive in practice. Goel et al. (2022) proposes EUk method, which retrains only the last k layers of the model while freezing the other layers. As a result, it is computationally cheaper than retraining the entire model. They also propose the CFk method, which continues training the last k layers on the DkeepsubscriptkeepD_keepDkeep without retraining from scratch, while freezing the other layers. • Gradient Ascent (GA) (Golatkar et al., 2020a; Graves et al., 2021; Jang et al., 2023): Gradient ascent is arguably the most popular heuristic for machine unlearning. It seeks to remove the influence of the DforgetsubscriptforgetD_forgetDforget from the trained model by reversing the gradient updates associated with DforgetsubscriptforgetD_forgetDforget. Notably, researchers have reported that gradient ascent can lead to significant model utility degradation in some cases (Ilharco et al., 2023; Pawelczyk et al., 2024a). • NegGrad+++ (Kurmanji et al., 2024): NegGrad+++ is a combination of gradient ascent on DforgetsubscriptforgetD_forgetDforget and gradient descent on DkeepsubscriptkeepD_keepDkeep. It finetunes the current model by optimizing: β⋅^x∼Dkeep[ℓ(M;x)]−(1−β)^x∼Dforget[ℓ(M;x)]⋅subscript^similar-tosubscriptkeepdelimited-[]ℓ1subscript^similar-tosubscriptforgetdelimited-[]ℓ β· E_x D_keep[ (M;x)]-(1-% β) E_x D_forget[ (M;x)]β ⋅ over start_ARG blackboard_E end_ARGx ∼ D start_POSTSUBSCRIPT keep end_POSTSUBSCRIPT [ ℓ ( M ; x ) ] - ( 1 - β ) over start_ARG blackboard_E end_ARGx ∼ D start_POSTSUBSCRIPT forget end_POSTSUBSCRIPT [ ℓ ( M ; x ) ] where β∈(0,1)01β∈(0,1)β ∈ ( 0 , 1 ) is a hyperparameter and ^ Eover start_ARG blackboard_E end_ARG is the empirical expectation. The intuition is to “review” the information from DkeepsubscriptkeepD_keepDkeep in order to prevent the model degradation due to the gradient ascent. • SCRUB (Kurmanji et al., 2024): SCalable Remembering and Unlearning unBound (SCRUB) is a state-of-the-art unlearning method that leverages a student-teacher framework. It updates the model by optimizing the objective: ^x∼Dkeep[KL(Mlearn(x)∥M(x))+ℓ(M;x)]subscript^similar-tosubscriptkeepdelimited-[]KLconditionalsubscriptlearnℓ E_x D_keep[KL(M_% learn(x)\|M(x))+ (M;x)]over start_ARG blackboard_E end_ARGx ∼ D start_POSTSUBSCRIPT keep end_POSTSUBSCRIPT [ KL ( Mlearn ( x ) ∥ M ( x ) ) + ℓ ( M ; x ) ] −^x∼Dforget[KL(Mlearn(x)∥M(x))]subscript^similar-tosubscriptforgetdelimited-[]KLconditionalsubscriptlearn - E_x D_forget[KL(M_% learn(x)\|M(x))]- over start_ARG blackboard_E end_ARGx ∼ D start_POSTSUBSCRIPT forget end_POSTSUBSCRIPT [ KL ( Mlearn ( x ) ∥ M ( x ) ) ] where KL is the Kullback-Leibler divergence. SCRUB shares a similar intuition with NegGrad+++, which can also be viewed as a combination of gradient ascent on DforgetsubscriptforgetD_forgetDforget and descent on DkeepsubscriptkeepD_keepDkeep. Nevertheless, instead of directly employing the original loss ℓ ℓ, SCRUB leverages the KL divergence to the original model MlearnsubscriptlearnM_learnMlearn. It provides a different regularization compared to NegGrad+++. • Langevin Unlearning (Chien et al., 2024a, b): Langevin Unlearning leverages noisy gradient descent for machine unlearning. Specifically, during the training process, it replaces the common gradient descent with DP-SGD (Abadi et al., 2016). For unlearning process, it finetunes the model on DkeepsubscriptkeepD_keepDkeep with DP-SGD as well. Chien et al. (2024a) establishes a smooth theoretical connection between differential privacy and unlearning and shows that Langevin Unlearning can provide a formal privacy guarantee for non-convex problems. Unfortunately, they mentioned that the resulting privacy bound is too loose to be applied in practice. We test Langevin Unlearning empirically in our experiments. All the above unlearning methods fall under approximate unlearning (Thudi et al., 2022), valued for their practical efficiency. In contrast, exact unlearning methods, such as the sharding-based framework SISA (Bourtoule et al., 2021), demand significant computational and storage resources. SISA achieves exact unlearning by training multiple models independently on disjoint data partitions, but this deviates from standard machine learning workflows and introduces considerable memory overhead. 5.1 Enforcing the Same Computation Budget for Unlearning Methods We categorize all methods as follows: those that only require the forget set (RL, GA), those that only require the keep set (EUk, CFk, Langevin), and those that require both the forget and keep sets (NegGrad+, SCRUB). Since machine unlearning is about the trade-off between privacy-utility-efficiency (Guo et al., 2020; Chien et al., 2024a; Liu et al., 2024c), we carefully ensure a similar computational complexity for all tested unlearning methods when demonstrating the privacy-utility trade-off. We define a Complexity Unit as the gradient computation budget of one training epoch on |Dforget|subscriptforget|D_forget|| Dforget | samples and limit all unlearning methods to a maximum of 10 Complexity Units. Since |Dforget|=Usubscriptforget|D_forget|=U| Dforget | = U is roughly 1%percent11\%1 % of |Dtrain|subscripttrain|D_train|| Dtrain | throughout our experiments, all unlearning methods are indeed much more efficient than retraining from scratch (Pawelczyk et al., 2024a). For unlearning approaches that leverage DforgetsubscriptforgetD_forgetDforget only, they can be tuned via unlearning process for at most 10 epochs. For those that leverage DkeepsubscriptkeepD_keepDkeep only, we randomly subsample it to size U for each epoch and unlearn for at most 10 epochs. For methods that leverage both DforgetsubscriptforgetD_forgetDforget and DkeepsubscriptkeepD_keepDkeep simultaneously, we limit their maximum unlearning epoch to 5. The situation is slightly more complicated for EUk and CFk approaches since only the last k layers are trained to save computation. We randomly select U/rU/rU / r samples from the keep set in each epoch, where r is the ratio of trainable parameters in the last k layers compared to the total number of parameters in the model. Our setup ensures that all tested unlearning approaches exhibit a similar unlearning computational complexity for a fair comparison. We optimize the unlearning epoch for each method under the 10 Complexity Unit constraint by the following criterion: if the perplexity of the unlearned model on DtrainsubscripttrainD_trainDtrain increases by more than 1 point compared to that of the initial model (No Unlearn), we stop at the first epoch where this condition is met; otherwise, we use the checkpoint from the last epoch. 6 Experiments Table 2: The privacy leakage (PL) for each unlearning method against different attackers for GPT-2 / Llama-2 7B on the Enron-Phone dataset. The number in the parenthesis is the excess ratio of PL magnitude for cases Canary and Minority compared to Random, where a larger PL magnitude implies a more severe underestimation of privacy leakage in the standard unlearning evaluation (Random). Bold font indicates the case that the amount of underestimated privacy leakage is at least 20%. Method PL (lossMIA) PL (zlibMIA) PL (Min-K%) Random Canary Minority Random Canary Minority Random Canary Minority Enron-Phone Dataset / GPT-2 No Unlearn 0.190 0.283 (49%↑ ↑) 0.340 (79%↑ ↑) 0.052 0.076 (48%↑ ↑) 0.064 (24%↑ ↑) 0.300 0.447 (49%↑ ↑) 0.524 (75%↑ ↑) RL 0.118 0.191 (61%↑ ↑) 0.210 (77%↑ ↑) 0.044 0.067 (52%↑ ↑) 0.060 (37%↑ ↑) 0.258 0.401 (55%↑ ↑) 0.447 (73%↑ ↑) EUk 0.027 0.080 (198%↑ ↑) 0.124 (362%↑ ↑) 0.035 0.051 (47%↑ ↑) 0.052 (49%↑ ↑) 0.092 0.215 (134%↑ ↑) 0.223 (143%↑ ↑) CFk 0.190 0.278 (46%↑ ↑) 0.337 (77%↑ ↑) 0.053 0.075 (41%↑ ↑) 0.064 (21%↑ ↑) 0.298 0.435 (46%↑ ↑) 0.514 (73%↑ ↑) GA 0.089 0.140 (57%↑ ↑) 0.127 (42%↑ ↑) 0.024 0.042 (73%↑ ↑) 0.026 (7%↑ ↑) 0.151 0.242 (60%↑ ↑) 0.171 (13%↑ ↑) NegGrad+ 0.183 0.271 (48%↑ ↑) 0.327 (79%↑ ↑) 0.052 0.073 (42%↑ ↑) 0.058 (13%↑ ↑) 0.293 0.435 (48%↑ ↑) 0.511 (74%↑ ↑) SCRUB 0.167 0.251 (50%↑ ↑) 0.321 (92%↑ ↑) 0.048 0.070 (44%↑ ↑) 0.062 (28%↑ ↑) 0.295 0.450 (52%↑ ↑) 0.527 (78%↑ ↑) Langevin 0.093 0.144 (54%↑ ↑) 0.157 (69%↑ ↑) 0.024 0.037 (54%↑ ↑) 0.027 (12%↑ ↑) 0.160 0.258 (61%↑ ↑) 0.264 (65%↑ ↑) Enron-Phone Dataset / Llama-2 7B No Unlearn 0.060 0.242 (303%↑ ↑) 0.172 (187%↑ ↑) 0.034 0.098 (188%↑ ↑) 0.067 (97%↑ ↑) 0.076 0.115 (51%↑ ↑) 0.179 (136%↑ ↑) RL -0.242 -0.084 (65%↓ ↓) -0.055 (77%↓ ↓) -0.005 0.065 (1400%↑ ↑) 0.102 (2140%↑ ↑) -0.123 -0.073 (41%↓ ↓) 0.012 (90%↓ ↓) EUk 0.057 0.246 (332%↑ ↑) 0.185 (225%↑ ↑) 0.039 0.106 (172%↑ ↑) 0.082 (110%↑ ↑) 0.063 0.132 (110%↑ ↑) 0.189 (200%↑ ↑) CFk 0.057 0.236 (314%↑ ↑) 0.168 (195%↑ ↑) 0.032 0.094 (194%↑ ↑) 0.063 (97%↑ ↑) 0.072 0.108 (50%↑ ↑) 0.171 (138%↑ ↑) GA -0.562 -0.430 (23%↓ ↓) -0.464 (17%↓ ↓) -0.014 0.038 (371%↑ ↑) 0.083 (593%↑ ↑) -0.625 -0.459 (27%↓ ↓) -0.517 (17%↓ ↓) NegGrad+ -0.074 -0.184 (149%↑ ↑) -0.040 (46%↓ ↓) -0.021 -0.048 (129%↑ ↑) -0.002 (90%↓ ↓) -0.069 -0.271 (293%↑ ↑) -0.057 (17%↓ ↓) SCRUB 0.059 0.162 (175%↑ ↑) 0.170 (188%↑ ↑) 0.034 0.063 (85%↑ ↑) 0.065 (91%↑ ↑) 0.074 -0.031 (58%↓ ↓) 0.177 (139%↑ ↑) Langevin 0.033 0.180 (445%↑ ↑) 0.104 (215%↑ ↑) 0.016 0.068 (325%↑ ↑) 0.036 (125%↑ ↑) 0.033 0.055 (67%↑ ↑) 0.091 (176%↑ ↑) Datasets. Our LLM unlearning evaluation is conducted on two representative PII datasets: Enron (Klimt & Yang, 2004) and ECHR (Chalkidis et al., 2019). The Enron dataset contains corporate emails released by the Federal Energy Regulatory Commission, while the ECHR dataset comprises legal case information from the European Court of Human Rights. For our experiments, we focus on specific PIIs based on their distributions: phone numbers (Enron-Phone) and email domains (Enron-Email) in Enron, and the year of judgment (ECHR-Year) in ECHR. Data minorities are defined based on these PIIs. Detailed dataset statistics are provided in App. A.1. Our study centers on instance-level unlearning, treating each individual as a single record. General Settings. We focus on the fine-tuning scenario, where the initial model M0subscript0M_0M0 is a pretrained LLM (GPT-2 (Radford et al., 2019) or Llama-2 7B (Touvron et al., 2023)). The fine-tuned model MlearnsubscriptlearnM_learnMlearn is obtained by training M0subscript0M_0M0 on a dataset DtrainsubscripttrainD_trainDtrain for 5 epochs. In the GPT-2 experiments, both the training and test sets contain 10,000 samples, subsampled from the full dataset. For Llama-2, we employ efficient fine-tuning using LoRA (Hu et al., 2021); both the training and test sets consist of 50,000 samples, subsampled from the entire dataset. In all cases, the forget set size is set to 1% of the training set size. The models are optimized using the AdamW optimizer with a constant learning rate of 10−5superscript10510^-510- 5, following the settings described in Shi et al. (2024b). During the unlearning process, all unlearning methods are constrained to the same computational budget—not exceeding 10 complexity units—as detailed in Section 5.1. We ensure that the unlearning complexity of each method is similar to allow for a fair comparison. Our ultimate goal is to achieve a superior privacy-utility-efficiency trade-off. We utilize MIA to estimate the empirical privacy risk measured by the PL metric as described in Section 3.1. For evaluating the utility of the LLMs, we report the perplexity following standard practices in the literature (Radford et al., 2019; Zhang et al., 2022), where a lower perplexity indicates that the model is more confident in its predictions. Additional details are provided in App. B. 6.1 Standard Approaches Underestimate Privacy Risk for Minorities. We report the results pertaining to the Enron-Phone, Enron-Email, and the ECHR-Year datasets. The experiment setting follows the explanation in Section 4 and further details are relegated to App. A.1. Table 2 shows that across all three attackers (lossMIA, zlibMIA, and Min-K%), all six unlearning methods and original model (no unlearning), the privacy leakage measure is significantly larger when unlearning canaries and minorities on Enron-Phone dataset for GPT-2 and Llama-2 7B respectively. Notably, in almost all cases the privacy leakage is underestimated for at least 20%. A similar phenomenon holds for the Enron-Email and ECHR-Year (Table 4, 5, 6, 7 in App. C.1) datasets. These results verify our claim that the current LLM unlearning evaluation indeed understated the privacy risk, especially for minorities. Our results call for a more careful empirical LLM unlearning evaluation, where considering canaries and minorities as we described can be an effective first step. 6.2 Benchmarking Unlearning Approaches under Minority-Aware Evaluation. Figure 4: Benchmarking unlearning approaches via our minority-aware evaluation for GPT-2 and Llama-2 on Enron-Phone (Left) and Enron-Email (Right) datasets. (a),(c),(e),(g): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for Min-K% attack. (b),(d),(f),(h): Worst perplexity over the three cases of each method. More results on lossMIA and zlibMIA attackers are deferred to App. C.2. Motivated by our observations, we propose the minority-aware LLM unlearning evaluation. Instead of reporting the privacy leakage (PL) score under the Random case, we propose to report the magnitude of maximum PL score of three settings (Random, Canary, and Minority) . This provides a better privacy risk estimation while keeping the entire evaluation pipeline efficient. Besides, we report the corresponding worst-case perplexity as the utility measure for each unlearning approach. We benchmark the popular unlearning methods under our new evaluation pipeline, where the result is summarized in Figure 4 for GPT-2 and Llama-2 on the Enron-Phone and Enron-Email datasets. See App. C.2 for additional results. We found that Langevin Unlearning offers the best balance between privacy and utility empirically. Note that while gradient ascent has on-par performance compared to Langevin Unlearning on the Enron-Phone dataset, it significantly degrades the model utility on the Enron-Email dataset. This echoes the finding of Ilharco et al. (2023); Pawelczyk et al. (2024a), albeit for different tasks. We found that gradient ascent is inherently unstable. In contrast, unlearning methods that leverage keep set DkeepsubscriptkeepD_keepDkeep are much more stable, including Langevin Unlearning and SCRUB. We also present results using utility metrics, including BERTScore (Zhang et al., 2019) and ROUGE (Lin, 2004), which capture semantic meaning. These results, provided in App. C.3, exhibit a consistent trend. 6.3 Ablation studies. We present ablation studies on the Enron-Phone dataset using the GPT-2 model, unless otherwise specified, and further results are deferred to App. C.4, C.5 and C.6. The effect of unlearning epochs for each unlearning approach.(a): Maximum PL with the attacker being Min-K%. (b): Model perplexity. The effect of forget set size for each unlearning approach.(a): Maximum PL with the attacker being Min-K%. (Results on lossMIA, zlibMIA are deferred to App. C.5). (b): Model perplexity. Figure 5: Ablation studies on unlearning iterations and forget set size. Privacy-Utility Trade-off. We analyze the privacy-utility trade-off curves for stable methods like Langevin Unlearning and SCRUB, along with the widely used GA. For Langevin Unlearning, we adjust the noise scale during training and unlearning. In SCRUB, we vary the weights that balance the loss and KL regularizer terms in its objective function (Sec. 5). For GA, we explore different learning rates ranging from 1e−7171e-71 e - 7 to 1e−3131e-31 e - 3. As shown in Fig.6, Figure 6: Privacy-utility Trade-off Curves for GPT-2. these curves are evaluated on the Enron-Phone and Enron-Email datasets. The results indicate that Langevin Unlearning (Green Line) outperforms SCRUB (Purple Line) with a superior privacy-utility trade-off. Notably, while GA performs reasonably well on the Enron-Phone dataset, its trade-off on Enron-Email is significantly weaker, revealing its instability. Detailed hyperparameter tuning for these methods is provided in App.C.6. Unlearning Iteration. We investigate the effect of unlearning epochs on privacy (Max PL) and utility (perplexity) for each unlearning approach in Fig. 5 (Left). We observe that RL and GA are unstable in PL score. Furthermore, these two methods can lead to significant model utility degradation in terms of perplexity, where even unlearning for 2 epochs can already result in a model breakdown. This observation again demonstrates that gradient ascent, albeit being simple and popular, is not a reliable LLM unlearning solution. We should focus on stable unlearning solutions such as SCRUB and Langevin Unlearning. Size of Forget Set. In Fig. 5 (Right), we report the effect of different forget set sizes on the privacy (Max PL) and utility (Perplexity) trade-offs for each unlearning method. We find that both the GA and RL methods are highly sensitive to the forget set size, leading to significant model utility degradation and poor reliability in practice. In contrast, methods like Langevin Unlearning demonstrate good performance in terms of stability. 7 Conclusions We identify a critical limitation in the typical evaluation pipeline for LLM unlearning efficacy: privacy risks to minority groups in the training data are often underestimated. Through carefully designed experiments using unlearning canaries tied to minority groups, inspired by privacy auditing research, we demonstrate that minority groups face at least 20% greater privacy leakage on average. Using personally identifiable information (PII) as a proxy for minority identifiers, we emphasize the need for more rigorous evaluations to ensure the right to be forgotten applies universally. Benchmarking existing unlearning methods with our minority-aware evaluation reveals that popular heuristics like gradient ascent are unstable and can degrade model utility. In contrast, methods such as Langevin Unlearning achieve a more favorable privacy-utility trade-off. Impact Statement This work introduces a minority-aware evaluation framework aimed at enhancing the comprehensiveness and equitable nature of LLM unlearning assessments. Our research contributes to the standardization of user data protection and the responsible utilization of machine learning data. However, we emphasize the importance of using compliant, publicly available, or officially authorized data when conducting minority-aware evaluations with LLMs. Building on these foundations, our work aims to foster more ethical and inclusive practices in the development and application of machine learning technologies. Acknowledgement This work is supported by NSF awards CCF-2402816, IIS-2239565, and JPMC faculty award 2023. The authors would like to thank Junyuan Hong and Chulin Xie for their assistance in reproducing the code for LLM-PBE: Assessing Data Privacy in Large Language Models. We are also grateful to Martin Pawelczyk and Ayush Sekhari for their valuable clarifications on their work, and to Peter Kairouz for his insightful comments. Disclaimer This paper was prepared for informational purposes in part by the Artificial Intelligence Research group of JPMorgan Chase & Co. and its affiliates (”JP Morgan”) and is not a product of the Research Department of JP Morgan. JP Morgan makes no representation and warranty whatsoever and disclaims all liability, for the completeness, accuracy or reliability of the information contained herein. This document is not intended as investment research or investment advice, or a recommendation, offer or solicitation for the purchase or sale of any security, financial instrument, financial product or service, or to be used in any way for evaluating the merits of participating in any transaction, and shall not constitute a solicitation under any jurisdiction or to any person, if such solicitation under such jurisdiction or to such person would be unlawful. References Abadi et al. (2016) Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., and Zhang, L. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, CCS ’16, p. 308–318, New York, NY, USA, 2016. Association for Computing Machinery. ISBN 9781450341394. doi: 10.1145/2976749.2978318. URL https://doi.org/10.1145/2976749.2978318. Aerni et al. (2024) Aerni, M., Zhang, J., and Tramèr, F. Evaluations of machine learning privacy defenses are misleading. In ACM SIGSAC Conference on Computer and Communications Security (CCS), 2024. URL https://arxiv.org/abs/2404.17399. Bourtoule et al. (2021) Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C. A., Jia, H., Travers, A., Zhang, B., Lie, D., and Papernot, N. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), p. 141–159. IEEE, 2021. Cao & Yang (2015) Cao, Y. and Yang, J. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, p. 463–480. IEEE, 2015. Carlini et al. (2019) Carlini, N., Liu, C., Erlingsson, Ú., Kos, J., and Song, D. The secret sharer: Evaluating and testing unintended memorization in neural networks. In 28th USENIX security symposium (USENIX security 19), p. 267–284, 2019. Carlini et al. (2021) Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), p. 2633–2650, 2021. Carlini et al. (2022) Carlini, N., Chien, S., Nasr, M., Song, S., Terzis, A., and Tramer, F. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), p. 1897–1914. IEEE, 2022. Chalkidis et al. (2019) Chalkidis, I., Androutsopoulos, I., and Aletras, N. Neural legal judgment prediction in english. arXiv preprint arXiv:1906.02059, 2019. Chen & Yang (2023) Chen, J. and Yang, D. Unlearn what you want to forget: Efficient unlearning for LLMs. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. URL https://openreview.net/forum?id=GprvtTwOxy. Chien et al. (2023) Chien, E., Pan, C., and Milenkovic, O. Efficient model updates for approximate unlearning of graph-structured data. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=fhcu4FBLciL. Chien et al. (2024a) Chien, E., Wang, H., Chen, Z., and Li, P. Langevin unlearning: A new perspective of noisy gradient descent for machine unlearning. arXiv preprint arXiv:2401.10371, 2024a. Chien et al. (2024b) Chien, E., Wang, H. P., Chen, Z., and Li, P. Certified machine unlearning via noisy stochastic gradient descent. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024b. URL https://openreview.net/forum?id=h3k2NXu5bJ. Duan et al. (2024) Duan, M., Suri, A., Mireshghallah, N., Min, S., Shi, W., Zettlemoyer, L., Tsvetkov, Y., Choi, Y., Evans, D., and Hajishirzi, H. Do membership inference attacks work on large language models? arXiv preprint arXiv:2402.07841, 2024. Feldman & Zhang (2020) Feldman, V. and Zhang, C. What neural networks memorize and why: Discovering the long tail via influence estimation. Advances in Neural Information Processing Systems, 33:2881–2891, 2020. Goel et al. (2022) Goel, S., Prabhu, A., Sanyal, A., Lim, S.-N., Torr, P., and Kumaraguru, P. Towards adversarial evaluations for inexact machine unlearning. arXiv preprint arXiv:2201.06640, 2022. Golatkar et al. (2020a) Golatkar, A., Achille, A., and Soatto, S. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 9304–9312, 2020a. Golatkar et al. (2020b) Golatkar, A., Achille, A., and Soatto, S. Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX 16, p. 383–398. Springer, 2020b. Graves et al. (2021) Graves, L., Nagisetty, V., and Ganesh, V. Amnesiac machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, p. 11516–11524, 2021. Guo et al. (2020) Guo, C., Goldstein, T., Hannun, A., and Van Der Maaten, L. Certified data removal from machine learning models. In I, H. D. and Singh, A. (eds.), Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, p. 3832–3842. PMLR, 13–18 Jul 2020. URL https://proceedings.mlr.press/v119/guo20c.html. Hayes et al. (2024) Hayes, J., Shumailov, I., Triantafillou, E., Khalifa, A., and Papernot, N. Inexact unlearning needs more careful evaluations to avoid a false sense of privacy. arXiv preprint arXiv:2403.01218, 2024. Hu et al. (2021) Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. Ilharco et al. (2023) Ilharco, G., Ribeiro, M. T., Wortsman, M., Schmidt, L., Hajishirzi, H., and Farhadi, A. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=6t0Kwf8-jrj. Jagielski et al. (2020) Jagielski, M., Ullman, J., and Oprea, A. Auditing differentially private machine learning: How private is private sgd? Advances in Neural Information Processing Systems, 33:22205–22216, 2020. Jang et al. (2022) Jang, J., Yoon, D., Yang, S., Cha, S., Lee, M., Logeswaran, L., and Seo, M. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. Jang et al. (2023) Jang, J., Yoon, D., Yang, S., Cha, S., Lee, M., Logeswaran, L., and Seo, M. Knowledge unlearning for mitigating privacy risks in language models. In Rogers, A., Boyd-Graber, J., and Okazaki, N. (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 14389–14408, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.805. URL https://aclanthology.org/2023.acl-long.805. Klimt & Yang (2004) Klimt, B. and Yang, Y. Introducing the enron corpus. In CEAS, volume 4, p. 1, 2004. Krzysztofek (2018) Krzysztofek, M. GDPR: General Data Protection Regulation (EU) 2016/679: Post-reform Personal Data Protection in the European Union. Kluwer Law International BV, 2018. Kurmanji et al. (2024) Kurmanji, M., Triantafillou, P., Hayes, J., and Triantafillou, E. Towards unbounded machine unlearning. Advances in neural information processing systems, 36, 2024. Li et al. (2024a) Li, Q., Hong, J., Xie, C., Tan, J., Xin, R., Hou, J., Yin, X., Wang, Z., Hendrycks, D., Wang, Z., Li, B., He, B., and Song, D. Llm-pbe: Assessing data privacy in large language models. Proc. VLDB Endow., 17(11):3201–3214, August 2024a. ISSN 2150-8097. doi: 10.14778/3681954.3681994. URL https://doi.org/10.14778/3681954.3681994. Li et al. (2024b) Li, W., Li, J., de Witt, C. S., Prabhu, A., and Sanyal, A. Delta-influence: Unlearning poisons via influence functions. arXiv preprint arXiv:2411.13731, 2024b. Lin (2004) Lin, C.-Y. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, p. 74–81, 2004. Liu et al. (2024a) Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., SHARMA, P., Liu, S., et al. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36, 2024a. Liu et al. (2024b) Liu, S., Yao, Y., Jia, J., Casper, S., Baracaldo, N., Hase, P., Xu, X., Yao, Y., Li, H., Varshney, K. R., et al. Rethinking machine unlearning for large language models. arXiv preprint arXiv:2402.08787, 2024b. Liu et al. (2024c) Liu, Z., Dou, G., Chien, E., Zhang, C., Tian, Y., and Zhu, Z. Breaking the trilemma of privacy, utility, and efficiency via controllable machine unlearning. In Proceedings of the ACM on Web Conference 2024, p. 1260–1271, 2024c. Lukas et al. (2023) Lukas, N., Salem, A., Sim, R., Tople, S., Wutschitz, L., and Zanella-Béguelin, S. Analyzing leakage of personally identifiable information in language models. In 2023 IEEE Symposium on Security and Privacy (SP), p. 346–363. IEEE, 2023. Maini et al. (2024) Maini, P., Feng, Z., Schwarzschild, A., Lipton, Z. C., and Kolter, J. Z. Tofu: A task of fictitious unlearning for llms. arXiv preprint arXiv:2401.06121, 2024. Nasr et al. (2021) Nasr, M., Songi, S., Thakurta, A., Papernot, N., and Carlin, N. Adversary instantiation: Lower bounds for differentially private machine learning. In 2021 IEEE Symposium on security and privacy (SP), p. 866–882. IEEE, 2021. Nasr et al. (2023) Nasr, M., Hayes, J., Steinke, T., Balle, B., Tramèr, F., Jagielski, M., Carlini, N., and Terzis, A. Tight auditing of differentially private machine learning. In 32nd USENIX Security Symposium (USENIX Security 23), p. 1631–1648, 2023. Neel et al. (2021) Neel, S., Roth, A., and Sharifi-Malvajerdi, S. Descent-to-delete: Gradient-based methods for machine unlearning. In Algorithmic Learning Theory, p. 931–962. PMLR, 2021. Ouyang et al. (2022) Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022. Pawelczyk et al. (2024a) Pawelczyk, M., Di, J. Z., Lu, Y., Kamath, G., Sekhari, A., and Neel, S. Machine unlearning fails to remove data poisoning attacks. arXiv preprint arXiv:2406.17216, 2024a. Pawelczyk et al. (2024b) Pawelczyk, M., Neel, S., and Lakkaraju, H. In-context unlearning: Language models as few-shot unlearners. In Forty-first International Conference on Machine Learning, 2024b. URL https://openreview.net/forum?id=GKcwle8XC9. Radford et al. (2019) Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. Sekhari et al. (2021) Sekhari, A., Acharya, J., Kamath, G., and Suresh, A. T. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075–18086, 2021. Shi et al. (2023) Shi, W., Ajith, A., Xia, M., Huang, Y., Liu, D., Blevins, T., Chen, D., and Zettlemoyer, L. Detecting pretraining data from large language models. arXiv preprint arXiv:2310.16789, 2023. Shi et al. (2024a) Shi, W., Ajith, A., Xia, M., Huang, Y., Liu, D., Blevins, T., Chen, D., and Zettlemoyer, L. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations, 2024a. URL https://openreview.net/forum?id=zWqr3MQuNs. Shi et al. (2024b) Shi, W., Lee, J., Huang, Y., Malladi, S., Zhao, J., Holtzman, A., Liu, D., Zettlemoyer, L., Smith, N. A., and Zhang, C. Muse: Machine unlearning six-way evaluation for language models. arXiv preprint arXiv:2407.06460, 2024b. Shokri et al. (2017) Shokri, R., Stronati, M., Song, C., and Shmatikov, V. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), p. 3–18. IEEE, 2017. Steinke & Ullman (2020) Steinke, T. and Ullman, J. The pitfalls of average-case differential privacy. DifferentialPrivacy.org, 07 2020. https://differentialprivacy.org/average-case-dp/. Steinke et al. (2024) Steinke, T., Nasr, M., and Jagielski, M. Privacy auditing with one (1) training run. Advances in Neural Information Processing Systems, 36, 2024. Thudi et al. (2022) Thudi, A., Jia, H., Shumailov, I., and Papernot, N. On the necessity of auditable algorithmic definitions for machine unlearning. In 31st USENIX Security Symposium (USENIX Security 22), p. 4007–4022, 2022. Thudi et al. (2024) Thudi, A., Jia, H., Meehan, C., Shumailov, I., and Papernot, N. Gradients look alike: Sensitivity is often overestimated in \DP-SGD\. In 33rd USENIX Security Symposium (USENIX Security 24), p. 973–990, 2024. Touvron et al. (2023) Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. Ullah & Arora (2023) Ullah, E. and Arora, R. From adaptive query release to machine unlearning. In International Conference on Machine Learning, p. 34642–34667. PMLR, 2023. Ullah et al. (2021) Ullah, E., Mai, T., Rao, A., Rossi, R. A., and Arora, R. Machine unlearning via algorithmic stability. In Conference on Learning Theory, p. 4126–4142. PMLR, 2021. Wu et al. (2023) Wu, X., Li, J., Xu, M., Dong, W., Wu, S., Bian, C., and Xiong, D. Depn: Detecting and editing privacy neurons in pretrained language models. arXiv preprint arXiv:2310.20138, 2023. Yao et al. (2024) Yao, J., Chien, E., Du, M., Niu, X., Wang, T., Cheng, Z., and Yue, X. Machine unlearning of pre-trained large language models. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8403–8419, Bangkok, Thailand, August 2024. Association for Computational Linguistics. URL https://aclanthology.org/2024.acl-long.457. Yao et al. (2023) Yao, Y., Xu, X., and Liu, Y. Large language model unlearning. In Socially Responsible Language Modelling Research, 2023. URL https://openreview.net/forum?id=wKe6jE065x. Ye et al. (2022) Ye, J., Maddi, A., Murakonda, S. K., Bindschaedler, V., and Shokri, R. Enhanced membership inference attacks against machine learning models. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, p. 3093–3106, 2022. Yeom et al. (2018) Yeom, S., Giacomelli, I., Fredrikson, M., and Jha, S. Privacy risk in machine learning: Analyzing the connection to overfitting. In 2018 IEEE 31st computer security foundations symposium (CSF), p. 268–282. IEEE, 2018. Zhang et al. (2024) Zhang, R., Lin, L., Bai, Y., and Mei, S. Negative preference optimization: From catastrophic collapse to effective unlearning. arXiv preprint arXiv:2404.05868, 2024. Zhang et al. (2022) Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X. V., et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. Zhang et al. (2019) Zhang, T., Kishore, V., Wu, F., Weinberger, K. Q., and Artzi, Y. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. Zhao et al. (2024) Zhao, K., Kurmanji, M., Bărbulescu, G.-O., Triantafillou, E., and Triantafillou, P. What makes unlearning hard and what to do about it. arXiv preprint arXiv:2406.01257, 2024. Appendix A Dataset A.1 Dataset Details In this paper, we examine two representative PII datasets: Enron and ECHR, described as follows: • Enron (Klimt & Yang, 2004). The Enron dataset consists of 536,000 authentic emails from 158 employees of the Enron Corporation, made publicly available by the Federal Energy Regulatory Commission following an investigation. Each email typically includes the sending timestamp, sender and recipient information, a greeting, the main content, and a footer containing the sender’s personal details. • ECHR (Chalkidis et al., 2019). The ECHR dataset comprises case records from the European Court of Human Rights. Each record contains a series of factual lists that detail the specifics of a case. In our experiments, we further decompose these cases into individual facts, with each fact forming a distinct sample, averaging around 80 tokens in length. In total, the dataset includes around 118,000 samples. A.2 PII Selection As outlined in Sec. 4, we selected U.S. phone numbers from the Enron dataset based on a criterion aimed at analyzing privacy risks in minority groups. To ensure the PII distribution was imbalanced, reflecting both minority and majority groups, we additionally selected two standard PII types (Lukas et al., 2023): email addresses (Enron) and years (ECHR). The distributions of these PII counts are depicted in Fig. 7. Figure 7: Histogram of email addresses (Enron-Email) and years (ECHR-Year). A.3 Preprocessing and Dataset Split Dataset Preprocess. In our experiments, since the average token length of Enron samples is approximately 770, we controlled the token length of each fact to ensure the model could effectively memorize the samples. We randomly selected three coherent sentences from each sample, and if the sample contained specific PIIs of interest, we prioritized selecting sentences around them. We will keep the original samples for the ECHR dataset. Dataset Construction. We begin by searching the dataset for occurrences of specific PIIs and analyzing their distribution. To form the minority set used in our Minority setting, we select 100 samples containing the least frequent PIIs; this set serves as our forget set. In the Random setting, we construct the forget set by randomly selecting 100 samples containing PIIs. To create the canary set, we replace the PIIs in the forget set (Random setting) with the least frequent PII found in the dataset. From the remaining data, we randomly select samples to create the training and test sets. For experiments with GPT-2 (117M), we uniformly at random selected 10,000 samples each for both the training and test sets. For Llama-2 7B, we uniformly at random selected 50,000 samples for both the training and test sets. Appendix B Experimental Details B.1 Compute Configurations All experiments were conducted using 8 NVIDIA A100 GPUs (80GB) and 14 NVIDIA RTX 6000 Ada GPUs (48GB). B.2 Unlearning Experiment Setup Unlearning Algorithms. For all unlearning methods, we use a constant learning rate of 10−5superscript10510^-510- 5 and a batch size of 32, consistent with the fine-tuning stage. Note that some unlearning algorithms require additional hyperparameters. We follow the common designs from previous literature (Pawelczyk et al., 2024a) and detail the hyperparameter selection as follows: • EUK and CFK. In our experiments, we set the number of retrained layers to k=33k=3k = 3 for both GPT-2 and Llama-2 (LoRA) models. For GPT-2, the unfrozen trainable parameters account for approximately 16% of the total parameters, while for Llama-2 7B, the unfrozen parameters account for around 10%. • NegGrad+. As noted in the main text, the hyperparameter β balances samples between DforgetsubscriptforgetD_forgetDforget and DkeepsubscriptkeepD_keepDkeep. In these experiments, we set β=0.9990.999β=0.999β = 0.999. • SCRUB. In the SCRUB method, three hyperparameters are used to balance the loss function on the keep set and the KL regularizers on both the keep and forget sets. According to the definition of the objective function in Section 5, three terms are weighted sequentially by setting: α=0.9990.999α=0.999α = 0.999, β=11β=1β = 1, and γ=0.010.01γ=0.01γ = 0.01. • Langevin Unlearning. The Langevin Unlearning method leverages noisy gradient descent to unlearn samples from the forget set. In our experiments, we set the Gaussian noise scale to σ=5e−454σ=5e-4σ = 5 e - 4 (for GPT-2) and σ=5e−353σ=5e-3σ = 5 e - 3 (for Llama-2), and the clipping norm to 1. Unlearning Epoch Selection. As outlined in Section 5.1, all unlearning methods are constrained to a maximum of 10 complexity units, and the optimal epoch for each method is selected based on whether the perplexity of the unlearned model on DtrainsubscripttrainD_trainDtrain increases by more than 1 point. Under our computational budget, methods that only require the forget set (RL, GA) are run for 10 epochs, while methods requiring both the keep and forget sets (NegGrad+, SCRUB) are limited to 5 epochs, due to the equal-sized cycling between the two sets. For methods that only require the keep set (EUK, CFK, Langevin), we use 10 epochs, with varying sample sizes for EUK and CFK, as some model parameters remain frozen. The selected epoch for each method in each experiment is detailed in Table 3. Table 3: Epochs comparison between unlearning methods on GPT2 and LLaMA2 models. Unlearning Methods GPT-2 Llama-2 7B Enron ECHR Enron ECHR RL Epoch 1 Epoch 1 Epoch 1 Epoch 1 EUk Epoch 10 Epoch 10 Epoch 10 Epoch 10 CFk Epoch 10 Epoch 10 Epoch 10 Epoch 10 GA Epoch 1 Epoch 1 Epoch 1 Epoch 1 NegGrad+ Epoch 5 Epoch 5 Epoch 1 Epoch 5 SCRUB Epoch 5 Epoch 5 Epoch 5 Epoch 5 Langevin Epoch 10 Epoch 10 Epoch 10 Epoch 10 Attack Method Hyperparameters. We employed three attack methods in our evaluation pipeline. For lossMIA and zlibMIA, there are no hyperparameters to tune. The Min-K%percentK\%K % method is based on the observation that non-member examples tend to have more tokens with lower likelihoods compared to member examples. In this method, the hyperparameter K controls the selection of the bottom K%percentK\%K % of tokens in each sample based on their likelihoods. Following previous recommendations in (Duan et al., 2024; Shi et al., 2024a), we set K=2020K=20K = 20 in our experiments. Random Seed Selection. In all our experiments, we followed the common practice and fixed our random seed to be 42. Appendix C Additional Experimental Results In this section, we present supplementary experimental results to further substantiate our claims in the main text. C.1 Experiments on Enron-Email and ECHR-year datasets In Tables 4, 7, 5 and 7, we report the PL scores for all three attackers across the three scenarios on the Enron-email, ECHR-year datasets for GPT-2 and Llama-2, respectively. The results support our claim that the current LLM unlearning evaluation (Random setting) significantly underestimates privacy risk. Table 4: The privacy leakage (PL) for each unlearning method against different attackers for GPT-2 on the Enron-Email dataset. Method PL (lossMIA) PL (zlibMIA) PL (Min-K%) Random Canary Minority Random Canary Minority Random Canary Minority No Unlearn 0.303 0.535 (77%↑ ↑) 1.145 (278%↑ ↑) 0.200 0.309 (54%↑ ↑) 0.262 (31%↑ ↑) 0.529 0.934 (76%↑ ↑) 1.468 (178%↑ ↑) RL 0.033 0.153 (366%↑ ↑) 0.448 (1265%↑ ↑) 0.062 0.142 (127%↑ ↑) 0.121 (94%↑ ↑) 0.431 0.772 (79%↑ ↑) 1.200 (179%↑ ↑) EUk 0.232 0.440 (89%↑ ↑) 0.582 (150%↑ ↑) 0.135 0.229 (70%↑ ↑) 0.152 (13%↑ ↑) 0.501 0.886 (77%↑ ↑) 1.034 (106%↑ ↑) CFk 0.296 0.515 (74%↑ ↑) 1.139 (285%↑ ↑) 0.197 0.295 (49%↑ ↑) 0.260 (32%↑ ↑) 0.526 0.905 (72%↑ ↑) 1.478 (181%↑ ↑) GA -0.279 -0.173 (38%↓ ↓) 0.739 (165%↑ ↑) -0.119 -0.037 (69%↓ ↓) 0.168 (41%↑ ↑) -0.390 -0.304 (22%↓ ↓) 1.034 (165%↑ ↑) NegGrad+ 0.265 0.471 (77%↑ ↑) 1.103 (316%↑ ↑) 0.179 0.269 (50%↑ ↑) 0.251 (40%↑ ↑) 0.496 0.864 (74%↑ ↑) 1.434 (189%↑ ↑) SCRUB 0.286 0.499 (74%↑ ↑) 1.097 (283%↑ ↑) 0.190 0.289 (52%↑ ↑) 0.253 (34%↑ ↑) 0.519 0.902 (74%↑ ↑) 1.473 (184%↑ ↑) Langevin 0.154 0.319 (107%↑ ↑) 0.606 (293%↑ ↑) 0.086 0.178 (107%↑ ↑) 0.124 (44%↑ ↑) 0.336 0.645 (92%↑ ↑) 0.940 (180%↑ ↑) Table 5: The privacy leakage (PL) for each unlearning method against different attackers for GPT-2 on ECHR-year datasets. Method PL (lossMIA) PL (zlibMIA) PL (Min-K%) Random Canary Minority Random Canary Minority Random Canary Minority No Unlearn 0.198 0.247 (25%↑ ↑) 0.263 (33%↑ ↑) 0.086 0.103 (20%↑ ↑) 0.122 (42%↑ ↑) 0.213 0.276 (30%↑ ↑) 0.299 (40%↑ ↑) RL 0.161 0.213 (32%↑ ↑) 0.234 (45%↑ ↑) 0.067 0.088 (31%↑ ↑) 0.086 (28%↑ ↑) 0.190 0.259 (36%↑ ↑) 0.257 (35%↑ ↑) EUk 0.125 0.176 (41%↑ ↑) 0.138 (10%↑ ↑) 0.067 0.088 (31%↑ ↑) 0.070 (4%↑ ↑) 0.114 0.187 (64%↑ ↑) 0.135 (18%↑ ↑) CFk 0.188 0.234 (24%↑ ↑) 0.260 (38%↑ ↑) 0.084 0.095 (13%↑ ↑) 0.120 (43%↑ ↑) 0.209 0.264 (26%↑ ↑) 0.295 (41%↑ ↑) GA 0.067 0.027 (60%↓ ↓) 0.105 (57%↑ ↑) 0.024 0.019 (21%↓ ↓) 0.038 (58%↑ ↑) 0.090 -0.019 (79%↓ ↓) 0.143 (59%↑ ↑) NegGrad+ 0.183 0.221 (21%↑ ↑) 0.247 (35%↑ ↑) 0.071 0.088 (24%↑ ↑) 0.112 (58%↑ ↑) 0.191 0.237 (24%↑ ↑) 0.274 (43%↑ ↑) SCRUB 0.179 0.223 (25%↑ ↑) 0.253 (41%↑ ↑) 0.080 0.099 (24%↑ ↑) 0.116 (45%↑ ↑) 0.197 0.253 (38%↑ ↑) 0.289 (47%↑ ↑) Table 6: The privacy leakage (PL) for each unlearning method against different attackers for Llama-2 7B on the Enron-Email dataset. Method PL (lossMIA) PL (zlibMIA) PL (Min-K%) Random Canary Minority Random Canary Minority Random Canary Minority No Unlearn 0.050 0.282 (464%↑ ↑) 0.174 (248%↑ ↑) 0.046 0.236 (413%↑ ↑) 0.095 (106%↑ ↑) 0.064 0.474 (640%↑ ↑) 0.224 (250%↑ ↑) RL -0.609 -0.567 (7%↓ ↓) -0.821 (12%↓ ↓) -0.832 -0.874 (5%↑ ↑) -0.478 (43%↓ ↓) -0.931 -0.931 (0%) -0.849 (9%↓ ↓) EUk 0.037 0.241 (551%↑ ↑) 0.169 (356%↑ ↑) 0.015 0.186 (1140%↑ ↑) 0.102 (580%↑ ↑) 0.040 0.364 (810%↑ ↑) 0.206 (415%↑ ↑) CFk 0.049 0.264 (438%↑ ↑) 0.169 (245%↑ ↑) 0.046 0.220 (378%↑ ↑) 0.090 (96%↑ ↑) 0.062 0.436 (603%↑ ↑) 0.218 (251%↑ ↑) GA -0.512 -0.692 (35%↑ ↑) 0.059 (89%↓ ↓) -0.435 -0.232 (47%↓ ↓) 0.184 (58%↓ ↓) -0.569 -0.479 (16%↓ ↓) 0.294 (48%↓ ↓) NegGrad+ -0.931 -0.929 (2%↓ ↓) -0.821 (12%↓ ↓) -0.832 -0.874 (5%↑ ↑) -0.478 (43%↓ ↓) -0.931 -0.931 (0%) -0.849 (9%↓ ↓) SCRUB 0.040 0.257 (543%↑ ↑) 0.174 (335%↑ ↑) 0.034 0.209 (515%↑ ↑) 0.095 (179%↑ ↑) 0.056 0.426 (661%↑ ↑) 0.224 (300%↑ ↑) Langevin 0.022 0.191 (768%↑ ↑) 0.048 (118%↑ ↑) 0.020 0.141 (605%↑ ↑) 0.021 (5%↑ ↑) 0.035 0.339 (868%↑ ↑) 0.079 (126%↑ ↑) Table 7: The privacy leakage (PL) for each unlearning method against different attackers for Llama-2 7B on the ECHR-year dataset. Method PL (lossMIA) PL (zlibMIA) PL (Min-K%) Random Canary Minority Random Canary Minority Random Canary Minority No Unlearn 0.056 0.094 (68%↑ ↑) 0.076 (35%↑ ↑) 0.030 0.048 (60%↑ ↑) 0.096 (220%↑ ↑) 0.067 0.114 (70%↑ ↑) 0.138 (106%↑ ↑) RL -0.069 0.044 (36%↓ ↓) -0.532 (671%↑ ↑) -0.024 0.029 (21%↑ ↑) -0.192 (700%↑ ↑) -0.034 0.070 (106%↑ ↑) -0.458 (1247%↑ ↑) EUk 0.059 0.084 (42%↑ ↑) 0.079 (34%↑ ↑) 0.030 0.044 (47%↑ ↑) 0.079 (163%↑ ↑) 0.065 0.110 (69%↑ ↑) 0.153 (135%↑ ↑) CFk 0.056 0.088 (57%↑ ↑) 0.073 (30%↑ ↑) 0.028 0.044 (57%↑ ↑) 0.088 (214%↑ ↑) 0.063 0.106 (68%↑ ↑) 0.131 (108%↑ ↑) GA -0.046 -0.376 (717%↑ ↑) -0.624 (1257%↑ ↑) -0.016 -0.120 (650%↑ ↑) -0.267 (1569%↑ ↑) -0.063 -0.404 (541%↑ ↑) -0.574 (811%↑ ↑) NegGrad+ 0.024 -0.272 (1033%↑ ↑) -0.624 (2500%↑ ↑) 0.012 -0.099 (725%↑ ↑) -0.235 (1858%↑ ↑) 0.026 -0.404 (1454%↑ ↑) -0.663 (2449%↑ ↑) SCRUB 0.056 0.094 (68%↑ ↑) 0.073 (30%↑ ↑) 0.030 0.048 (60%↑ ↑) 0.090 (200%↑ ↑) 0.067 0.112 (67%↑ ↑) 0.139 (107%↑ ↑) Langevin 0.026 0.052 (100%↑ ↑) 0.041 (58%↑ ↑) 0.010 0.025 (150%↑ ↑) 0.046 (360%↑ ↑) 0.028 0.062 (121%↑ ↑) 0.078 (179%↑ ↑) C.2 More Results on Minority-Aware Evaluation In this section, we present further benchmarking results for unlearning approaches under minority-aware LLM evaluation. Following the same setup as Section 6.2, Fig. 8 reports the maximum PL score under lossMIA and zlibMIA attackers on Enron-Phone and Enron-Email and Fig. 9 reports the maximum PL score and worst-case perplexity for various unlearning methods on ECHR-Year (GPT-2) dataset. Figure 8: Benchmarking unlearning approaches via our minority-aware evaluation for GPT-2 and Llama-2 on Enron-Phone and Enron-Email dataset. (a),(c),(e),(g): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for lossMIA attack. (b),(d),(f),(h): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for zlibMIA attack. Figure 9: Benchmarking unlearning approaches via our minority-aware evaluation for GPT-2 on ECHR-year dataset. (a)-(c): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for lossMIA, zlibMIA, and Min-K% attacks respectively. (d): Worst perplexity over the three cases of each method. Figure 10: Benchmarking unlearning approaches via our minority-aware evaluation for Llama-2 on ECHR Year dataset. (a)-(c): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for lossMIA, zlibMIA, and Min-K% attacks respectively. (d): Worst perplexity over the three cases of each method. We observe that both GA and Langevin Unlearning methods maintain a favorable balance between privacy and utility. However, GA can be sensitive to the forget set size and the number of unlearning iterations (Section 6.3). In practice, the GA method should be applied with caution, whereas more stable approaches like Langevin Unlearning offer a better trade-off in terms of privacy, utility, and stability. C.3 Results on Other Utility Metrics In this section, we report the worst-case utility performance across three settings (Random, Canary, and Minority) using utility metrics BERTScore (Zhang et al., 2019) and ROUGE score (Lin, 2004), which capture semantic meaning, on the Enron-Email dataset. The results are shown in Fig.11 (GPT-2) and Fig.12 (Llama-2). As illustrated in the figures, the performance of Random Label and gradient-ascent-based methods (Gradient Ascent and NegGrad+) is unstable under all utility metrics (Perplexity, BERTScore, and ROUGE). In contrast, Langevin Unlearning demonstrates relatively stable performance and achieves a favorable privacy-utility trade-off. Figure 11: Benchmarking unlearning approaches via our minority-aware evaluation for GPT-2 on Enron-email dataset. (a)-(c): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for lossMIA, zlibMIA, and Min-K% attacks respectively. (d-f): Worst utility performance over the three cases of each method. Figure 12: Benchmarking unlearning approaches via our minority-aware evaluation for Llama-2 on Enron-email dataset. (a)-(c): Maximum privacy leakage (PL) over three cases (Random, Canary, and Minority) for lossMIA, zlibMIA, and Min-K% attacks respectively. (d-f): Worst utility performance over the three cases of each method. C.4 Further Details and Results on Langevin Unlearning In this section, we provide additional details and results on the Langevin Unlearning methods. As mentioned in Section 5, Langevin leverages noisy gradient descent and involves training the model on the dataset DtrainsubscripttrainD_trainDtrain using DP-SGD. Furthermore, Langevin conducts machine unlearning by fine-tuning the model on the dataset DkeepsubscriptkeepD_keepDkeep with DP-SGD as well. It is important to note that for the Langevin Unlearning method, the training process incorporates noise. Consequently, our retrain baseline is adjusted to train the initial model on DkeepsubscriptkeepD_keepDkeep using DP-SGD for 5 epochs. Furthermore, in Table 8 and 9, we report the effectiveness of Langevin Unlearning by evaluating it against three MIA methods (lossMIA, zlibMIA, and Min-K%) across different datasets on GPT-2. These evaluations are conducted across three scenarios (Random, Canary, Minority), assessing the PL scores, the maximum PL scores and the worst-case perplexity. By comparing the results of the Noisy No Unlearn baseline (which fine-tunes the initial model with DP-SGD for 5 epochs) with those of the Langevin Unlearning method, we observe that minority scenarios (Canary, Minority) lead to significantly higher privacy leakage, and Langevin Unlearning achieves superior privacy-utility trade-offs. Additionally, in practical applications, the number of steps employing noisy gradient descent can be tailored based on the acceptable computational costs, thereby enabling potentially better privacy-utility trade-offs. This flexibility allows practitioners to balance the trade-off between enhanced privacy and computational efficiency according to specific application requirements. Table 8: The privacy leakage (PL) for Langevin Unlearning against different attackers for GPT-2 on All datasets. Method PL (lossMIA) PL (zlibMIA) PL (Min-K%) Random Canary Minority Random Canary Minority Random Canary Minority Enron-phone Noisy No Unlearn 0.097 0.152 (57%↑ ↑) 0.170 (75%↑ ↑) 0.024 0.039 (63%↑ ↑) 0.033 (38%↑ ↑) 0.164 0.259 (58%↑ ↑) 0.268 (63%↑ ↑) Langevin 0.092 0.144 (57%↑ ↑) 0.157 (71%↑ ↑) 0.024 0.037 (54%↑ ↑) 0.027 (13%↑ ↑) 0.159 0.259 (63%↑ ↑) 0.264 (66%↑ ↑) Enron-email Noisy No Unlearn 0.156 0.342 (119%↑ ↑) 0.642 (312%↑ ↑) 0.102 0.193 (89%↑ ↑) 0.130 (27%↑ ↑) 0.344 0.691 (101%↑ ↑) 0.945 (175%↑ ↑) Langevin 0.154 0.319 (107%↑ ↑) 0.606 (294%↑ ↑) 0.097 0.178 (84%↑ ↑) 0.124 (28%↑ ↑) 0.336 0.645 (92%↑ ↑) 0.939 (179%↑ ↑) ECHR-year Noisy No Unlearn 0.101 0.152 (51%↑ ↑) 0.122 (21%↑ ↑) 0.049 0.067 (37%↑ ↑) 0.064 (31%↑ ↑) 0.122 0.180 (48%↑ ↑) 0.145 (19%↑ ↑) Langevin 0.103 0.140 (36%↑ ↑) 0.125 (21%↑ ↑) 0.049 0.061 (24%↑ ↑) 0.065 (33%↑ ↑) 0.117 0.168 (44%↑ ↑) 0.146 (25%↑ ↑) Table 9: Maximum PL Scores and Worst-case Perplexity for Noisy No Unlearn and Langevin across Datasets on GPT-2 Dataset Methods lossMIA zlibMIA Min-K% Perplexity Enron phone Noisy No Unlearn 0.170 0.039 0.268 13.87 Langevin 0.157 (7.65%↓ ↓) 0.037 (5.13%↓ ↓) 0.264 (1.49%↓ ↓) 13.88 Enron email Noisy No Unlearn 0.642 0.193 0.945 12.52 Langevin 0.606 (5.61%↓ ↓) 0.178 (7.77%↓ ↓) 0.939 (0.63%↓ ↓) 12.61 ECHR year Noisy No Unlearn 0.152 0.067 0.180 12.75 Langevin 0.140 (7.89%↓ ↓) 0.061 (8.96%↓ ↓) 0.168 (6.67%↓ ↓) 12.78 C.5 More Results on Forget Set Size In this section, we report additional results on the impact of forget set size for each unlearning method, using LossMIA and ZlibMIA attackers. As shown in Fig. 13, similar to the results in the main text, both RL and GA methods are sensitive to the forget set size, whereas methods like SCRUB and Langevin Unlearning demonstrate greater stability. Figure 13: The effect of forget set size for each unlearning approach. (a)(b): Maximum PL over three cases (Random, Canary, Minority) with the attacker being lossMIA and zlibMIA respectively. C.6 More Details on the Privacy-Utility Trade-off Curves for Langevin Unlearning and SCRUB Methods This section provides comprehensive experiments of the privacy-utility trade-off curves for the Langevin Unlearning and SCRUB methods, as introduced in Sec. 6.3. We detailed the hyperparameters used in the ablation study as follows: Figure 14: Privacy-utility transition curves for Enron-Phone dataset with hyperparameter β=11β=1β = 1 (Top) and hyperparameter β=1e−313β=1e-3β = 1 e - 3 (Bottom). Figure 15: Privacy-utility transition curves for Enron-Email dataset with hyperparameter β=11β=1β = 1 (Top) and hyperparameter β=1e−313β=1e-3β = 1 e - 3 (Bottom). Langevin Unlearning. For the Langevin Unlearning method, we fix the clipping norm to 1 and vary the noise scale added during training to control the privacy-utility trade-off. In experiments with the GPT-2 model, the noise scale σ is adjusted across the values 1e−4,3e−4,5e−4,8e−4,1e−31434548413\1e-4,3e-4,5e-4,8e-4,1e-3\ 1 e - 4 , 3 e - 4 , 5 e - 4 , 8 e - 4 , 1 e - 3 . Table 19 and 20 present the AUC scores under various attackers (lossMIA, zlibMIA, Min-k%) and utility (perplexity) on the Enron-Phone and Enron-Email datasets, respectively. SCRUB. The SCRUB training objective comprises the original loss ℓ ℓ on the keep set, along with two KL divergence regularizers on the keep and forget sets. These terms are balanced by three hyperparameters: ^x∼Dkeep[αKL(Mlearn(x)∥M(x))+βℓ(M;x)]−^x∼Dforget[γKL(Mlearn(x)∥M(x))].subscript^similar-tosubscriptkeepdelimited-[]KLconditionalsubscriptlearnℓsubscript^similar-tosubscriptforgetdelimited-[]KLconditionalsubscriptlearn E_x D_keep[ (M_ % learn(x)\|M(x))+β (M;x)]- E_x D_forget[% (M_learn(x)\|M(x))].over start_ARG blackboard_E end_ARGx ∼ D start_POSTSUBSCRIPT keep end_POSTSUBSCRIPT [ α KL ( Mlearn ( x ) ∥ M ( x ) ) + β ℓ ( M ; x ) ] - over start_ARG blackboard_E end_ARGx ∼ D start_POSTSUBSCRIPT forget end_POSTSUBSCRIPT [ γ KL ( Mlearn ( x ) ∥ M ( x ) ) ] . (2) We conducted an extensive hyperparameter search, setting β to 1 and 1e−3131e-31 e - 3 in separate configurations. For each fixed β, α and γ are independently varied from 1e−4,5e−4,1e−3,5e−3,1e−2,5e−2,1e−1,5e−1,114541353125211511\1e-4,5e-4,1e-3,5e-3,1e-2,5e-2,1e-1,5e-1,1\ 1 e - 4 , 5 e - 4 , 1 e - 3 , 5 e - 3 , 1 e - 2 , 5 e - 2 , 1 e - 1 , 5 e - 1 , 1 . Fig. 14 and 15 illustrates the resulting transition curves, showing the maximum privacy leakage (PL) for three scenarios (Random, Canary, Minority) across different attackers (lossMIA, zlibMIA, Min-k%) and utility (perplexity) metrics on both Enron-Phone and Enron-Email datasets. The transition curves highlight how SCRUB’s performance depends on balancing the three objective terms. Notably, when the KL regularizer weight on the keep set is greater than or equal to that on the forget set, SCRUB achieves relatively high utility, albeit with increased privacy leakage. Besides, we observe in our experiments that the zlibMIA attacker fails to capture the inherent privacy-utility trade-off for SCRUB as demonstrated in the transition curves. We further report the privacy-utility trade-off curves for both methods under attacker being lossMIA and zlibMIA. Similar to the results demonstrated in Sec. 6.3, Langevin Unlearning method achieves the best trade-off performance over SCRUB method. Figure 16: Privacy-utility trade-off curves under lossMIA and zlibMIA. Figure 17: Degree of the largest underestimation in privacy leakage (Canary, Minority) compared to Random settings across varying Complexity Units. C.7 Privacy Underestimation Across Complexity Units In Fig. 17, we report the degree of largest underestimation (Canary & Minoirty settings) in privacy leakage compared to Random settings across different complexity units under Min-k% attacker. As demonstrated in the figure, under different complexity units, the privacy leakage for each unlearning methods are severely underestimated. Detailed results under each settings are reported in Table 11 and 12. C.8 Discussion on Retraining LLMs with Equivalent Computational Budget In this subsection, we examine the scenario where the LLM is retrained from scratch using a computational budget equivalent to that of unlearning methods. Intuitively, if retraining with the same computational resources could already achieve a satisfactory privacy-utility trade-off, employing unlearning methods would be less important. To empirically illustrate this point, we consider the Enron (Phone) dataset with the GPT-2 model as a representative case study. Specifically, we report the privacy leakage metrics (PrivLeak) across different attack scenarios, as well as the corresponding perplexity scores in Tab. 10. Additionally, the performance of the retrained model with an equivalent computational budget is highlighted in Fig. 18 (Black). Our findings reveal that while retraining alone indeed achieves relatively low privacy leakage, it drastically compromises model utility. These results underscore the importance and practical necessity of developing and evaluating unlearning methods. Table 10: Privacy Leakage (PL) for Retrain with 10 Complexity Units. Random Canary Minority Loss PL 0.004 0.043 0.083 Zlib PL 0.058 0.075 0.077 Min-k% PL 0.083 0.174 0.188 Perplexity 23.44 Figure 18: Privacy-Utility Trade-off for Enron-Phone Min-k% with Retrain (Same Computation Budget). C.9 Results on AUC Scores, Perplexity across Different Models and Datasets We further report the AUC scores under different attackers (lossMIA, zlibMIA, Min-K%) and utility (perplexity) over holdout test set DtestsubscripttestD_testDtest for GPT-2 and Llama-2 in Table. 13-18. C.10 Discussion on TPR@low FPR Metric Note that aside from the AUC score, a commonly reported metric for privacy evaluation is TPR@low FPR (Carlini et al., 2022), where the low FPR is often set to 0.01. However, in our scenario, the canary size is set to 100 (1% of the total training set size). At FPR=0.01FPR0.01FPR=0.01FPR = 0.01, the TPR would be calculated based on only a few canary samples, making the overall score very coarse. To avoid the impact of this coarse granularity on our experimental results, we primarily focus on the AUC score. Table 11: PL Scores for MIA across Three Settings for GPT-2 on Enron (Phone) under Different Complexity Units. Enron-Phone GPT-2 Min-k% Methods Complexity Units 2 Complexity Units 4 Complexity Units 10 Random Canary Minority Random Canary Minority Random Canary Minority Random Label 0.166 0.284 0.136 -0.059 0.037 0.290 0.168 0.320 0.355 EUk 0.068 0.169 0.234 0.068 0.174 0.202 0.092 0.216 0.223 CFk 0.302 0.445 0.518 0.300 0.438 0.519 0.298 0.435 0.514 Gradient Ascent -0.175 -0.245 -0.203 -0.245 -0.469 -0.352 -0.031 -0.012 -0.427 NegGrad+ 0.309 0.452 0.543 0.309 0.447 0.529 0.298 0.452 0.510 Scrub 0.311 0.443 0.525 0.313 0.447 0.525 0.306 0.445 0.532 Langevin 0.163 0.259 0.265 0.161 0.259 0.265 0.159 0.259 0.265 Table 12: Perplexity for MIA across Three Settings for GPT-2 on Enron (Phone) under Different Complexity Units. Enron-Phone GPT-2 Perplexity Methods Complexity Units 2 Complexity Units 4 Complexity Units 10 Random Canary Minority Random Canary Minority Random Canary Minority Random Label 29.85 28.93 30+ 30+ 30+ 30+ 30+ 30+ 30+ EUk 30+ 30+ 30+ 30+ 30+ 30+ 23.64 23.65 26.60 CFk 12.72 12.71 12.72 12.72 12.72 12.73 12.67 12.67 12.67 Gradient Ascent 16.63 15.76 29.28 30+ 30+ 30+ 30+ 30+ 30+ NegGrad+ 12.86 12.87 12.83 12.84 12.86 12.88 12.83 12.80 12.88 Scrub 12.94 12.95 12.96 13.11 13.09 13.19 13.09 12.88 12.96 Langevin 13.84 13.85 13.88 13.85 13.86 13.88 13.84 13.88 13.88 Table 13: AUC Scores for MIA and Perplexity across Three Settings for GPT-2 on Enron (Phone Numbers) AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority No Unlearn 0.533 0.531 0.422 0.694 0.693 0.619 Retrain 0.448 0.414 0.315 0.660 0.644 0.582 Noisy No Unlearn 0.474 0.464 0.365 0.678 0.674 0.606 Noisy Retrain 0.432 0.403 0.312 0.662 0.649 0.587 Unlearning Methods Random Label 0.501 0.493 0.381 0.689 0.687 0.617 Langevin 0.472 0.461 0.361 0.678 0.673 0.603 EUk 0.460 0.447 0.354 0.683 0.677 0.612 CFk 0.533 0.529 0.421 0.695 0.692 0.619 Gradient Ascent 0.488 0.472 0.355 0.676 0.671 0.597 NegGrad+ 0.530 0.526 0.418 0.694 0.691 0.616 SCRUB 0.523 0.518 0.416 0.692 0.689 0.618 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority No Unlearn 0.594 0.592 0.471 12.72 12.72 12.72 Retrain 0.457 0.409 0.309 12.74 12.74 12.74 Noisy No Unlearn 0.518 0.511 0.393 13.84 13.85 13.87 Noisy Retrain 0.445 0.406 0.310 13.84 13.84 13.83 Unlearning Methods Random Label 0.575 0.573 0.447 14.49 14.41 14.86 Langevin 0.516 0.511 0.392 13.84 13.88 13.88 EUk 0.499 0.497 0.378 23.64 23.65 23.60 CFk 0.593 0.587 0.468 12.67 12.67 12.67 Gradient Ascent 0.526 0.508 0.362 13.22 13.20 14.10 NegGrad+ 0.591 0.587 0.467 12.86 12.86 12.88 SCRUB 0.592 0.593 0.472 13.00 12.98 12.96 Table 14: AUC Scores for MIA and Perplexity across Three Settings for GPT-2 on Enron (Email) AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority No Unlearn 0.555 0.551 0.354 0.462 0.462 0.563 Retrain 0.426 0.359 0.165 0.385 0.353 0.446 Noisy No Unlearn 0.488 0.483 0.271 0.422 0.421 0.512 Noisy Retrain 0.422 0.360 0.165 0.383 0.353 0.453 Unlearning Methods Random Label 0.440 0.414 0.239 0.409 0.403 0.500 Langevin 0.487 0.475 0.265 0.420 0.416 0.509 EUk 0.525 0.517 0.261 0.437 0.434 0.514 CFk 0.552 0.544 0.353 0.461 0.457 0.562 Gradient Ascent 0.307 0.297 0.287 0.339 0.340 0.521 NegGrad+ 0.539 0.528 0.347 0.454 0.448 0.558 SCRUB 0.548 0.538 0.346 0.458 0.455 0.559 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority No Unlearn 0.607 0.611 0.506 12.11 12.11 12.12 Retrain 0.397 0.316 0.205 12.19 12.19 12.19 Noisy No Unlearn 0.516 0.519 0.387 12.51 12.52 12.50 Noisy Retrain 0.384 0.307 0.199 12.48 12.48 12.48 Unlearning Methods Random Label 0.568 0.560 0.451 15.87 16.13 13.54 Langevin 0.513 0.505 0.386 12.58 12.58 12.61 EUk 0.596 0.596 0.417 23.58 23.70 23.58 CFk 0.606 0.602 0.508 12.14 12.15 12.15 Gradient Ascent 0.242 0.220 0.417 27.30 21.86 14.53 NegGrad+ 0.594 0.589 0.499 12.34 12.34 12.37 SCRUB 0.603 0.601 0.507 12.15 12.15 12.32 Table 15: AUC Scores for MIA and Perplexity across Three Settings for GPT-2 on ECHR (Year) AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority No Unlearn 0.661 0.650 0.658 0.532 0.524 0.560 Retrain 0.552 0.521 0.521 0.490 0.475 0.499 Noisy No Unlearn 0.600 0.592 0.581 0.514 0.509 0.535 Noisy Retrain 0.545 0.514 0.518 0.490 0.477 0.503 Unlearning Methods Random Label 0.641 0.632 0.643 0.523 0.517 0.542 Langevin 0.601 0.586 0.583 0.514 0.506 0.536 EUk 0.621 0.613 0.593 0.523 0.517 0.534 CFk 0.656 0.643 0.656 0.531 0.520 0.559 Gradient Ascent 0.589 0.535 0.576 0.502 0.484 0.518 NegGrad+ 0.653 0.636 0.650 0.525 0.517 0.555 SCRUB 0.651 0.637 0.653 0.529 0.522 0.557 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority No Unlearn 0.671 0.661 0.673 11.81 11.81 11.81 Retrain 0.553 0.518 0.518 11.82 11.82 11.82 Noisy No Unlearn 0.615 0.609 0.586 12.74 12.74 12.75 Noisy Retrain 0.548 0.516 0.512 12.73 12.73 12.73 Unlearning Methods Random Label 0.658 0.652 0.651 12.70 12.70 12.67 Langevin 0.612 0.603 0.587 12.78 12.78 12.78 EUk 0.616 0.615 0.588 22.04 22.06 22.00 CFk 0.669 0.655 0.671 11.78 11.78 11.79 Gradient Ascent 0.603 0.508 0.592 12.11 12.20 12.11 NegGrad+ 0.659 0.641 0.660 12.01 12.04 12.03 SCRUB 0.662 0.649 0.668 12.02 12.03 12.01 Table 16: AUC Scores for MIA and Perplexity across Three Settings for Llama-2 on Enron (Phone Number) AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority No Unlearn 0.614 0.606 0.551 0.578 0.571 0.575 Retrain 0.579 0.488 0.470 0.559 0.520 0.539 Noisy No Unlearn 0.605 0.598 0.539 0.578 0.572 0.576 Noisy Retrain 0.584 0.500 0.482 0.568 0.533 0.554 Unlearning Methods Random Label 0.439 0.447 0.444 0.556 0.554 0.594 Langevin 0.603 0.590 0.532 0.577 0.569 0.574 EUk 0.612 0.608 0.557 0.581 0.575 0.583 CFk 0.612 0.603 0.549 0.577 0.569 0.573 Gradient Ascent 0.253 0.278 0.252 0.551 0.540 0.584 NegGrad+ 0.536 0.398 0.451 0.547 0.495 0.538 SCRUB 0.613 0.567 0.550 0.578 0.553 0.574 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority No Unlearn 0.611 0.610 0.579 9.45 9.47 9.48 Retrain 0.568 0.547 0.491 9.48 9.48 9.48 Noisy No Unlearn 0.600 0.602 0.570 10.21 10.21 10.21 Noisy Retrain 0.579 0.565 0.516 10.19 10.19 10.19 Unlearning Methods Random Label 0.498 0.507 0.497 2124 2559 2445 Langevin 0.598 0.596 0.563 10.20 10.21 10.20 EUk 0.604 0.619 0.584 11.00 11.24 10.82 CFk 0.609 0.606 0.575 9.43 9.45 9.46 Gradient Ascent 0.213 0.296 0.237 4e9 8e9 2e9 NegGrad+ 0.529 0.399 0.463 9.82 9.90 9.85 SCRUB 0.610 0.530 0.578 9.45 9.48 9.48 Table 17: AUC Scores for MIA and Perplexity across Three Settings for Llama-2 on Enron (Email) AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority No Unlearn 0.628 0.613 0.418 0.548 0.539 0.451 Retrain 0.598 0.478 0.356 0.524 0.436 0.412 Noisy No Unlearn 0.594 0.572 0.393 0.515 0.502 0.443 Retrain 0.579 0.470 0.375 0.503 0.433 0.434 Unlearning Methods Random Label 0.234 0.207 0.394 0.333 0.333 0.458 Langevin 0.592 0.560 0.393 0.513 0.494 0.443 EUk 0.620 0.593 0.416 0.532 0.517 0.454 CFk 0.627 0.604 0.416 0.548 0.532 0.449 Gradient Ascent 0.292 0.147 0.377 0.296 0.335 0.488 NegGrad+ 0.041 0.034 0.064 0.088 0.055 0.215 SCRUB 0.622 0.601 0.418 0.542 0.527 0.451 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority No Unlearn 0.632 0.619 0.421 4.83 4.84 4.84 Retrain 0.594 0.420 0.344 4.84 4.84 4.84 Noisy No Unlearn 0.592 0.569 0.397 5.38 5.38 5.38 Noisy Retrain 0.570 0.410 0.368 5.39 5.39 5.39 Unlearning Methods Random Label 0.230 0.181 0.330 730 542 255 Langevin 0.590 0.549 0.397 5.36 5.36 5.37 EUk 0.618 0.573 0.415 5.42 5.53 5.37 CFk 0.631 0.603 0.419 4.83 4.83 4.83 Gradient Ascent 0.256 0.219 0.445 4e8 6e12 6e12 NegGrad+ 0.041 0.029 0.052 12.15 14.73 6.20 SCRUB 0.627 0.599 0.421 4.86 4.86 4.86 Table 18: AUC Scores for MIA and Perplexity across Three Settings for Llama-2 on ECHR (Year) AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority No Unlearn 0.570 0.547 0.726 0.513 0.499 0.513 Retrain 0.540 0.500 0.675 0.498 0.476 0.468 Noisy No Unlearn 0.556 0.532 0.721 0.504 0.491 0.510 Noisy Retrain 0.541 0.502 0.685 0.498 0.476 0.476 Unlearning Methods Random Label 0.503 0.522 0.316 0.486 0.490 0.378 Langevin 0.555 0.528 0.715 0.503 0.488 0.498 EUk 0.572 0.542 0.728 0.513 0.497 0.505 CFk 0.570 0.544 0.724 0.512 0.497 0.509 Gradient Ascent 0.515 0.312 0.254 0.490 0.419 0.343 NegGrad+ 0.553 0.364 0.254 0.504 0.429 0.358 SCRUB 0.570 0.547 0.724 0.513 0.499 0.510 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority No Unlearn 0.573 0.559 0.686 4.89 4.89 4.89 Retrain 0.537 0.502 0.603 4.89 4.89 4.89 Noisy No Unlearn 0.553 0.541 0.675 5.03 5.03 5.02 Noisy Retrain 0.537 0.504 0.616 5.02 5.02 5.02 Unlearning Methods Random Label 0.519 0.537 0.327 90 129 127.43 Langevin 0.552 0.535 0.664 5.03 5.03 5.03 EUk 0.572 0.557 0.695 5.27 5.21 5.32 CFk 0.571 0.555 0.682 4.87 4.88 4.87 Gradient Ascent 0.503 0.299 0.257 7.94 10.68 29.48 NegGrad+ 0.551 0.299 0.203 4.96 5.10 5.41 SCRUB 0.573 0.558 0.687 4.88 4.88 4.87 Table 19: AUC Scores for MIA and Perplexity across Three Settings for GPT-2 on Enron (Phone Number) for Noisy Learning AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority Noisy No Unlearn 1e-4 0.541 0.536 0.427 0.699 0.697 0.621 Noisy Retrain 1e-4 0.446 0.412 0.309 0.658 0.643 0.577 Langevin 1e-4 0.540 0.535 0.424 0.698 0.695 0.619 Noisy No Unlearn 3e-4 0.492 0.484 0.381 0.682 0.678 0.607 Noisy Retrain 3e-4 0.436 0.405 0.311 0.660 0.647 0.583 Langevin 3e-4 0.492 0.483 0.379 0.682 0.678 0.606 Noisy No Unlearn 5e-4 0.474 0.464 0.365 0.678 0.674 0.606 Noisy Retrain 5e-4 0.432 0.403 0.312 0.662 0.649 0.587 Langevin 5e-4 0.472 0.461 0.361 0.678 0.673 0.603 Noisy No Unlearn 8e-4 0.463 0.450 0.352 0.677 0.672 0.606 Noisy Retrain 8e-4 0.428 0.401 0.311 0.664 0.653 0.590 Langevin 8e-4 0.460 0.448 0.350 0.676 0.670 0.604 Noisy No Unlearn 1e-3 0.459 0.446 0.348 0.677 0.671 0.606 Noisy Retrain 1e-3 0.427 0.400 0.311 0.665 0.655 0.592 Langevin 1e-3 0.456 0.443 0.347 0.675 0.670 0.604 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority Noisy No Unlearn 1e-4 0.599 0.599 0.465 12.26 12.24 12.27 Noisy Retrain 1e-4 0.454 0.410 0.303 12.25 12.25 12.25 Langevin 1e-4 0.598 0.596 0.460 12.32 12.32 12.33 Noisy No Unlearn 3e-4 0.538 0.535 0.409 13.20 13.20 13.21 Noisy Retrain 3e-4 0.447 0.405 0.308 13.19 13.19 13.19 Langevin 3e-4 0.538 0.535 0.408 13.24 13.24 13.22 Noisy No Unlearn 5e-4 0.518 0.511 0.393 13.84 13.85 13.87 Noisy Retrain 5e-4 0.445 0.406 0.310 13.84 13.84 13.83 Langevin 5e-4 0.516 0.511 0.392 13.84 13.88 13.88 Noisy No Unlearn 8e-4 0.505 0.498 0.378 14.40 14.39 14.39 Noisy Retrain 8e-4 0.442 0.405 0.313 14.42 14.42 14.42 Langevin 8e-4 0.502 0.497 0.376 14.44 14.44 14.43 Noisy No Unlearn 1e-3 0.500 0.492 0.374 14.71 14.70 14.70 Noisy Retrain 1e-3 0.440 0.408 0.313 14.73 14.73 14.73 Langevin 1e-3 0.497 0.487 0.371 14.74 14.74 14.73 Table 20: AUC Scores for MIA and Perplexity across Three Settings for GPT-2 on Enron (Phone Email) for Noisy Learning AUC - LossMIA AUC - ZlibMIA Random Canary Minority Random Canary Minority Noisy No Unlearn 1e-4 0.584 0.587 0.369 0.480 0.483 0.578 Noisy Retrain 1e-4 0.431 0.362 0.165 0.387 0.356 0.447 Langevin 1e-4 0.584 0.578 0.368 0.479 0.476 0.579 Noisy No Unlearn 3e-4 0.512 0.511 0.311 0.436 0.436 0.535 Noisy Retrain 3e-4 0.424 0.360 0.164 0.384 0.353 0.449 Langevin 3e-4 0.510 0.499 0.309 0.432 0.429 0.534 Noisy No Unlearn 5e-4 0.488 0.483 0.271 0.422 0.421 0.512 Noisy Retrain 5e-4 0.422 0.360 0.165 0.383 0.353 0.453 Langevin 5e-4 0.487 0.475 0.265 0.420 0.416 0.509 Noisy No Unlearn 8e-4 0.470 0.464 0.248 0.414 0.412 0.503 Noisy Retrain 8e-4 0.417 0.357 0.170 0.383 0.354 0.459 Langevin 8e-4 0.471 0.457 0.243 0.411 0.406 0.500 Noisy No Unlearn 1e-3 0.463 0.456 0.243 0.410 0.408 0.500 Noisy Retrain 1e-3 0.414 0.354 0.172 0.381 0.355 0.462 Langevin 1e-3 0.462 0.450 0.237 0.408 0.403 0.499 AUC - Min-K% Perplexity Random Canary Minority Random Canary Minority Noisy No Unlearn 1e-4 0.651 0.661 0.528 12.18 12.25 12.16 Noisy Retrain 1e-4 0.407 0.318 0.210 12.25 12.25 12.25 Langevin 1e-4 0.655 0.655 0.532 12.27 12.28 12.27 Noisy No Unlearn 3e-4 0.553 0.562 0.420 12.31 12.34 12.28 Noisy Retrain 3e-4 0.390 0.311 0.199 12.23 12.23 12.23 Langevin 3e-4 0.552 0.547 0.421 12.38 12.39 12.38 Noisy No Unlearn 5e-4 0.516 0.519 0.387 12.51 12.52 12.50 Noisy Retrain 5e-4 0.384 0.307 0.199 12.48 12.48 12.48 Langevin 5e-4 0.513 0.505 0.386 12.58 12.58 12.61 Noisy No Unlearn 8e-4 0.478 0.483 0.355 12.83 12.80 12.82 Noisy Retrain 8e-4 0.369 0.298 0.200 12.84 12.84 12.84 Langevin 8e-4 0.477 0.469 0.349 12.87 12.86 12.89 Noisy No Unlearn 1e-3 0.465 0.467 0.341 13.01 12.98 13.01 Noisy Retrain 1e-3 0.365 0.296 0.201 13.02 13.02 13.02 Langevin 1e-3 0.462 0.453 0.335 13.04 13.03 13.06