Paper deep dive
Traceback Translators Against Forgetting in Continual Fake Speech Detection
Enrico Gottardis, Mattia Tamiazzo, Simone Milani
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 7/15/2026, 6:37:09 AM
Summary
This paper proposes a continual learning approach for audio deepfake detection that mitigates catastrophic forgetting by integrating a lightweight traceback domain translator into a frozen ResNet18 classifier. The translator remaps new feature spaces to the original domain using CORAL and Prototype Consistency losses, preserving accuracy on previous datasets while adapting to new ones, including cross-lingual scenarios. Experimental results demonstrate high detection rates with minimal computational overhead compared to traditional retraining.
Entities (10)
Relation Signals (9)
Traceback Translator Network → mitigates → Catastrophic Forgetting
confidence 95% · we propose a forgetting-resilient solution based on the adoption of domain translators within a frozen detector... Experimental results show that this strategy enables the achievement of high detection rates... while minimizing the computational effort and preserving the detection accuracy on previous data.
ResNet18 → usedfor → Audio Deepfake Detection
confidence 95% · The classification model is based on a customized ResNet18 [7] (see Figure 1), where the first convolutional layer is frozen and the classifier head consists of two fully-connected layers with normalization and dropout.
ASVspoof 2019 → usedfor → Audio Deepfake Detection
confidence 95% · The dataset used to train the baseline model is the training subset of ASVspoof 2019.
FakeOrReal → usedfor → Audio Deepfake Detection
confidence 95% · FakeOrReal (FoR) [21]. The dataset contains English bona fide and synthetic speech generated by various Text-To-Speech (TTS) systems
In-the-Wild → usedfor → Audio Deepfake Detection
confidence 95% · This corpus aims to assess speech detectors in a real-world scenario and includes real and synthetic audio samples collected from English-speaking celebrities and politicians.
ADD 2022 → usedfor → Audio Deepfake Detection
confidence 95% · The dataset includes genuine and synthetic Mandarin Chinese audio samples. It is designed to evaluate deepfake detection systems across various conditions
→ →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Fake speech detectors are increasingly challenged by the development of new and more accurate generative models. To cope with this problem, continual learning techniques are nowadays widely considered feasible strategies for updating models to new datasets, but they also lead to decreased performance on previously seen samples (catastrophic forgetting). In this work, we propose a forgetting-resilient solution based on the adoption of domain translators within a frozen detector, which remaps the new feature spaces into the original ones by means of a traceback translator network. Experimental results show that this strategy enables the achievement of high detection rates with respect to traditional retraining, while minimizing the computational effort and preserving the detection accuracy on previous data.
Tags
Links
- Source: https://arxiv.org/abs/2607.12569v1
- Canonical: https://arxiv.org/abs/2607.12569v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
28,864 characters extracted from source content.
Expand or collapse full text
Traceback Translators Against Forgetting in Continual Fake Speech Detection †thanks: This work was funded by the European Union under the Italian National Recovery and Resilience Plan (NRRP) of NextGenerationEU, with a partnership on “Telecommunications of the Future” (PE00000001 - program “RESTART”). We thank the partnership and collaboration with the São Paulo Research Foundation (Fapesp) Horus project, Grant #2023/12865-8. This manuscript reflects only the authors’ views and opinions, neither EU nor the European Commission nor FAPESP can be considered responsible for them. All of the authors have revised the document and confirm the findings. Enrico Gottardis Mattia Tamiazzo Simone Milani Abstract Fake speech detectors are increasingly challenged by the development of new and more accurate generative models. To cope with this problem, continual learning techniques are nowadays widely considered feasible strategies for updating models to new datasets, but they also lead to decreased performance on previously seen samples (catastrophic forgetting). In this work, we propose a forgetting-resilient solution based on the adoption of domain translators within a frozen detector, which remaps the new feature spaces into the original ones by means of a traceback translator network. Experimental results show that this strategy enables the achievement of high detection rates with respect to traditional retraining, while minimizing the computational effort and preserving the detection accuracy on previous data. I Introduction and related works In recent years, the widespread adoption and development of synthetic speech generators [2] have forced researchers and cybersecurity experts to create new deepfake detectors [12]. As a matter of fact, existing detectors need to be updated through continual retraining in which old datasets are integrated with new samples, but unfortunately the access to the original data is not always guaranteed. A possible alternative consists in fine-tuning pre-trained models on the sole novel dataset, but this leads to undesired learning drifts, such as catastrophic forgetting, where knowledge learned from old samples is diluted by new data. To address such issues, transfer learning, domain adaptation and continual learning techniques have been recently investigated in the literature [27, 1, 3, 29, 4, 17, 14, 19]. Methods based on transfer learning and domain adaptation aim at adapting pre-trained models and transferring knowledge across domains. These approaches include fine-tuning using Low-Rank Adaptation [27], cross-lingual knowledge transfer [1], and unsupervised domain adaptation to handle distribution shifts [3]. Strategies based on continual learning employ memory-based and regularization methods, such as self-adaptive learning to modify the gradient direction and retain previously learned knowledge [29], and model distillation techniques [4, 17]. The authors of [14] propose a continual learning method to incrementally train fake audio detectors on new spoofing attacks. The model includes a knowledge distillation loss and a similarity constraint that aligns genuine sample representations across different domains, improving detection performance. To improve generalization across unseen data, a Mixture of Experts architecture is proposed in [19]. Multiple specialized experts are trained on specific datasets, and their predictions are combined through a gating mechanism. By distributing specialized knowledge across multiple modules, the architecture is able to outperform single-model baselines in a cross-dataset scenario. The study in [23] investigates the effectiveness of continual learning by comparing different retraining strategies. The detector is divided into two logical sections (encoder and classification modules), and experiments evaluate the classification performance by applying a continual learning approach [14] to the entire model, or selectively to one section while freezing the other. Experimental results show that retraining the encoder while freezing the classifier offers a better balance between learning novel data and maintaining previously acquired knowledge. In this paper, we introduce a lightweight traceback domain translator into a frozen classifier in order to adapt it to new deepfake generators and sampling setups and mitigate catastrophic forgetting. It is possible to summarize the main contributions as follows: • We propose a new continual learning approach for synthetic speech detection, where traditional fine-tuning is replaced by the introduction of a lightweight domain translator within a frozen classifier. • The translator-based approach is compared to other domain-adaptation solutions; the number of retrained parameters is minimized while maximizing accuracies on both new and old datasets. • The proposed solution was tested in a multilingual setup (including both English and Chinese sentences), showing that the proposed approach can also be used for a cross-lingual adaptation. • Performance was further improved by means of generative data augmentation based on diffusion models. I Continual learning and catastrophic forgetting in synthetic speech detection models In order to tackle the problem of forgetting in continual learning for synthetic audio classification, we tested different adaptation methods measuring their impacts on a simple but effective classifier. Figure 1: Architecture of the customized ResNet18 used in the work. I-A Dataset pre-processing The adopted architecture processes the input signal and generates sets of MFCC spectrograms including 128128-frequency coefficients and spanning a time window of 4242 frames around salient instants (leading to 128×42128× 42 matrices). Salient instants are selected at points where the signal evolves from a silenced interval to a voiced one since their relevance has already been highlighted in some previous works [15, 16]. Although multiple spectrograms can be extracted from a single file, for some classes the number of available samples was limited, leading to a high class imbalance between real and synthetic audio samples. To address this issue, we resorted to a dataset generative process based on a diffusion model implemented in [10]. Specifically, we employed a 2D U-Net architecture [22] as the core denoising component of the diffusion process, enabling the generation of synthetic spectrogram samples while preserving the spatial structure and dimensions of the input representations. To further compensate for the class imbalance, classic data augmentation was used, and an imbalanced sampler was adopted. I-B Audio deepfake detection The classification model is based on a customized ResNet18 [7] (see Figure 1), where the first convolutional layer is frozen and the classifier head consists of two fully-connected layers with normalization and dropout. The model was trained using a weighted custom Cross-Entropy loss with label smoothing, to reduce over-confidence: ℒCE=−1N∑i=1N∑c=0K−1wcy~i,clogp^i,c, _CE=- 1N _i=1^N _c=0^K-1w_c\, y_i,c\, p_i,c, (1) y~i,c=1−ϵ,c=yiϵ/K,c≠yi 31.2982pt y_i,c= where =[6,1,1,1,1,1,1]∈ℝKw=[6,1,1,1,1,1,1] ^K is the vector of weights, with wcw_c the c-th element of w, ϵ=0.08ε=0.08, K=7K=7 represents the number of classes, and N is the batch size. From experimental results, we noticed a tendency of the model to misclassify real samples as if they were synthetic from class 66. For this reason, we added an extra penalty defined as: ℒpenalty=λ∑i=1N(yi=0∧argmaxcp^i,c=6),L_penalty=λ _i=1^N1 (y_i=0\ \ _c p_i,c=6 ), (2) with λ=5.0λ=5.0, leading to a combined classifier loss: ℒclassifier=ℒCE+ℒpenalty.L_classifier=L_CE+L_penalty. (3) For validation, we employed a standard Cross-Entropy loss. I-C Continual Learning Strategies After obtaining a solid backbone model to perform detection on audio deepfakes, we analyzed different approaches to perform continual learning on other datasets. I-C1 Full retraining The first approach was to retrain the whole network on the new dataset. This common method is largely implemented by other researchers [23], leading to impressive performance. However, the main drawback is the catastrophic forgetting phenomenon: the model learns the new representation of data, forgetting the previous task. I-C2 Domain adaptation In order to reduce the catastrophic forgetting issue, we opted to retrain only part of the whole architecture. In particular, we considered adopting a domain adaptation approach by selectively retraining only the Batch Normalization (BN) layers, while keeping the rest of the model frozen (sometimes called Continual Normalization [20]). This method has two main advantages: the retraining of the model requires only the new dataset, on which the network will be fine-tuned, and the training will be faster, as the only learnable parameters are the BN layers. I-C3 Domain Translation This idea consists of shifting the distributions of real and fake samples in the target domain to make them similar to those in the source domain. Although the proposed translator shares similarities with adapter modules explored in [8, 11, 5], its role and characteristics are different. The translator is applied after the embedding layer and is trained to perform class-conditional alignment of real or fake feature distributions across domains. By freezing the backbone and adapting only this lightweight module, the approach limits catastrophic forgetting. To the best of our knowledge, the use of simple adapter-like architectures has not been previously explored for distribution alignment in audio deepfake detection. We considered the features created from the model after the embedding, which provides vectors in a 128128-dimensional latent space. Before providing these vectors to the fully connected part of the model, we processed them with the architecture described in Fig. 2, where a fully-connected model with a bottleneck db=32d_b=32 is followed by a residual connection =+f(LN())y=x+f(LN(x)), shown in the picture with a dashed line. Figure 2: Block diagram for the translator model. To promote the translation process, the training loss consists of three elements: the classifier loss described in (3), the CORrelation ALignment (CORAL) loss [24] and Prototype Consistency (PC) [30] loss. While the CORAL loss aims at minimizing the mean and variance between real and fake samples from the source and the target datasets, PC minimizes the intra-class distance. Given source features zs∈ℝNB×128z_s ^N_B× 128 with labels ys∈0,…,6y_s∈\0,…,6\ and target features zt∈ℝNB×128z_t ^N_B× 128 with labels yt∈0,1y_t∈\0,1\, with NBN_B being the batch size, we define their batch-wise means μsreal _s^real, μtreal _t^real, μsfake _s^fake, μtfake _t^fake as class prototypes and the corresponding covariance matrices as Σs _s and Σt _t. Let us define the norm ∥μs−μt∥22=(‖μsreal−μtreal‖22+‖μsfake−μtfake‖22) _s- _t _2^2= (\| _s^real- _t^real\|_2^2+\| _s^fake- _t^fake\|_2^2 ). The prototype consistency loss is defined as: ℒPC=12∥μs−μt∥22.L_PC= 12 _s- _t _2^2. (4) Denoting ∥⋅∥F2 · _F^2 as the squared matrix Frobenius norm, it is then possible to compute the CORAL loss as: ℒCORAL=∥μs−μt∥22⏟mean alignment+∥Σs−Σt∥F2⏟covariance alignment.L_CORAL= _s- _t _2^2_mean alignment+ _s- _t _F^2_covariance alignment. (5) The final training loss is defined as: ℒtotal=ℒclassifier+λCORALℒCORAL+λPCℒPCL_total=L_classifier+ _CORAL\,L_CORAL+ _PC\,L_PC (6) where λCORAL _CORAL and λPC _PC are two regularization terms, experimentally set to 0.050.05 and 0.010.01 respectively. I Experimental setup The experimental setting for our continual learning approach was defined based on several state-of-the-art corpora of samples. ASVspoof 2019 (ASV19) [26]. The dataset was introduced to assess countermeasures against synthetic and voice-converted speech in English and is organized into three subsets: train, development, and evaluation. It is one of the most widely used benchmark datasets for audio deepfake detection. FakeOrReal (FoR) [21]. The dataset contains English bona fide and synthetic speech generated by various Text-To-Speech (TTS) systems, including both open-source and commercial tools. The real speech content is sourced from publicly available speech datasets. In-The-Wild (ITW) [18]. This corpus aims to assess speech detectors in a real-world scenario and includes real and synthetic audio samples collected from English-speaking celebrities and politicians. Covering diverse recording conditions and generative methods, it is particularly useful for evaluating model robustness. ADD 2022 (ADD22) [28]. The dataset includes genuine and synthetic Mandarin Chinese audio samples. It is designed to evaluate deepfake detection systems across various conditions, such as low-quality and partially fake audio tracks. The dataset used to train the baseline model is the training subset of ASVspoof 2019. Although training was performed on individual spectrograms, inference results were aggregated at the audio-excerpt level using majority voting over multiple spectrograms extracted from the same sample. IV Results This section presents a preliminary backbone selection study, followed by the evaluation of the proposed continual learning strategies, and an ablation analysis. IV-A Backbone Architecture Selection Before evaluating the proposed continual learning strategies, we performed a preliminary study to identify a suitable backbone model for the detection task. Several architectures were considered, evaluating them on the ASVspoof 2019 development set. Table I reports the comparison among the investigated architectures. Among the evaluated models, the customized ResNet18 provided the best performance even compared to more complex architectures, like EfficientNet [25], MobileNet [9], and ConvNeXT-Tiny [13]. The Audio Spectrogram Transformer (AST) [6] achieved comparable accuracy; however, it requires substantially higher computational resources, with approximately 86 million trainable parameters compared to the 11 million parameters of ResNet18, as well as longer training times. TABLE I: Model comparison for classification task on ASVspoof 2019. Architecture Dev. accuracy Dev. loss Tr. epochs Custom ResNet18 0.994 0.112 240 AST [6] 0.994 0.193 324 ConvNeXT-Tiny [13] 0.904 0.329 296 ResNet18 - multiclass [7] 0.883 0.269 148 ResNet18 - binary [7] 0.860 0.398 110 EfficientNet B0 [25] 0.823 0.461 60 EfficientNet B1 [25] 0.776 0.536 30 MobileNet V3 Large [9] 0.753 0.588 100 TABLE I: Experimental results of the proposed methods. Results in brackets are referred to the performance on the ASVspoof 2019 evaluation set after specialization (retraining). The method is compared with the LCNN-based model in [23]. Method ASV19 FOR ITW ADD22 AUC (%) EER (%) AUC (%) EER (%) AUC (%) EER (%) AUC (%) EER (%) Baseline model 95.0 9.74 49.8 50.2 49.6 50.3 30.8 62.9 Full retraining – – 99.7 (61.2) 3.27 (43.2) 100 (59.6) 0.28 (43.8) 100 (22.1) 0.78 (71.8) Domain adaptation – – 97.4 (63.1) 8.26 (40.7) 96.3 (77.6) 9.15 (28.8) 99.3 (46.0) 4.02 (52.2) Domain translation – – 96.0 (95.0) 11.37 (9.74) 95.4 (95.0) 10.64 (9.74) 98.1 (95.0) 6.96 (9.74) CL ALL [23] 95.7 11.1 99.0 (94.0) 4.20 (13.6) 99.8 (92.1) 1.60 (15.4) – – CL ℳcM_c [23] – – 98.8 (93.2) 5.00 (14.0) 99.5 (93.1) 2.80 (14.6) – – CL ℳeM_e [23] – – 98.9 (93.2) 4.90 (13.4) 99.0 (91.8) 4.10 (15.1) – – (a) ROC plots for the full retraining of the model. (b) ROC plots for the domain adaptation approach. (c) ROC plots for the domain translation approach. Figure 3: Comparison of the ROC plots obtained testing the proposed methods. TABLE I: Number of parameters of our methods compared with the LCNN model employed in [23]. Method # Trained parameters Full retraining 11095K Domain adaptation 10K Domain translation 21K CL ALL [23] 5556K CL ℳcM_c [23] 521K CL ℳeM_e [23] 5035K IV-B Main Experimental Results The results obtained on the evaluation datasets are summarized in Table I. The baseline model achieves performance comparable to that reported in [23], evaluated using the Receiver Operating Characteristic (ROC) curve, and the Equal Error Rate (EER). Performance is also quantified using the Area Under the Curve (AUC). The results of the three continual learning strategies introduced in I-C are illustrated in Fig.3. It can be observed that full retraining performs very well on the target dataset (average 1.44%1.44\% EER on new dataset), but results in a dramatic loss of accuracy on the source dataset ASVspoof 2019 (−52%-52\% AUC and +52.9%+52.9\% EER on average). The largest loss is reasonably observed on ADD since the language change from English to Chinese. Performing a partial update (domain adaptation) proved to be more robust, as the average AUC decrease is 2.3%2.3\% on the target domain, but the loss on the source domain is 38%38\% AUC and EER increases by 30.8%30.8\% on average. It is also worth noting that only a small number of parameters must be retrained (see Table I), making it very efficient. The domain translation strategy represents a good trade-off between classification performances and computational complexity since the average AUC loss on target domains is 3.4%3.4\%, but we preserve the accuracy on the source domain (95.4%95.4\% AUC, 9.74%9.74\% EER) while updating only a limited number of parameters. The experimental results also show that the proposed approaches can be applied to cross-lingual adaptation, since ADD dataset includes Chinese language speech samples. The traceback translation model proves to be capable of handling datasets in different languages while keeping the same classification accuracy. IV-C Ablation studies To determine the impact of diffusion-based data augmentation on model performance, we trained the baseline model on only the ASVspoof 2019 training set, and then we included both original and generated samples. Experimental results show consistent performance gains with data augmentation: classification accuracy increased from 81.9%81.9\% to 83.7%83.7\%, while the AUC improved from 91.5%91.5\% to 95.0%95.0\%. Moreover, we enabled/disabled the different components in (6) and evaluated their impact on the final accuracy. With only ℒPCL_PC (Prototype-Consistency), EER for FoR/ITW/ADD increases to 10.75%10.75\%, 11.87%11.87\%, 7.34%7.34\% respectively, while using ℒCORALL_CORAL only (CORAL loss), the respective EER is 11.84%11.84\%, 9.05%9.05\%, 7.72%7.72\%, showing that their combination is, on average, more effective on all the datasets. V Conclusions In this paper, we investigated the effectiveness of multiple continual learning techniques for audio deepfake detection to mitigate the problem of catastrophic forgetting. Our experiments showed that all the approaches achieve good detection performance, but the full retraining and domain adaptation methods are not able to retain previously learned knowledge. We proposed a novel approach based on domain translation that achieves high detection rates without requiring the retraining of the backbone architecture and, at the same time, preserves the detection accuracy on the original dataset. Future work will focus on extending the evaluation to a larger number of datasets to test scalability and robustness against different types of domain shifts. Additionally, exploring different model architectures could further improve performance in cross-domain and cross-lingual audio deepfake detection. Acknowledgment The authors would like to thank Davide Salvi, Viola Negroni, Luca Bondi, Paolo Bestagini, and Stefano Tubaro of Politecnico di Milano for their support in the experimental comparison with [23]. References [1] Z. Ba, Q. Wen, P. Cheng, Y. Wang, F. Lin, L. Lu, and Z. Liu (2023) Transferring audio deepfake detection capability across languages. In Proceedings of the ACM Web Conference 2023, W ’23, New York, NY, USA, p. 2033–2044. External Links: ISBN 9781450394161, Link, Document Cited by: §I. [2] K. Bhagtani, A. K. S. Yadav, P. Bestagini, and E. J. Delp (2024) Are recent deepfake speech generators detectable?. In Proceedings of the 2024 ACM Workshop on Information Hiding and Multimedia Security, IH&MMSec ’24, New York, NY, USA, p. 277–282. External Links: ISBN 9798400706370, Link, Document Cited by: §I. [3] X. Chen, W. Lu, R. Zhang, J. Xu, X. Lu, L. Zhang, and J. Wei (2025) Continual unsupervised domain adaptation for audio deepfake detection. In 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , Hyderabad, India, p. 1–5. External Links: Document Cited by: §I. [4] F. Dong, Q. Tang, Y. Bai, and Z. Wang (2024) Advancing continual learning for robust deepfake audio classification. In TENCON 2024 - 2024 IEEE Region 10 Conference (TENCON), Vol. , Singapore, p. 302–305. External Links: Document Cited by: §I. [5] S. V. Eeckt and H. Van Hamme (2023) Using adapters to overcome catastrophic forgetting in end-to-end automatic speech recognition. In 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 1–5. External Links: Document Cited by: §I-C3. [6] Y. Gong, Y. Chung, and J. Glass (2021) AST: Audio Spectrogram Transformer. In Interspeech 2021, p. 571–575. External Links: Document, ISSN 2958-1796 Cited by: §IV-A, TABLE I. [7] K. He, X. Zhang, S. Ren, and J. Sun (2016-06) Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, p. . Cited by: §I-B, TABLE I, TABLE I. [8] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly (2019-09–15 Jun) Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, K. Chaudhuri and R. Salakhutdinov (Eds.), Proceedings of Machine Learning Research, Vol. 97, p. 2790–2799. External Links: Link Cited by: §I-C3. [9] A. Howard, M. Sandler, G. Chu, L. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan, Q. V. Le, and H. Adam (2019-10) Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Cited by: §IV-A, TABLE I. [10] Hugging Face API models UNet2DModel. Note: Accessed on 2025-09-10 External Links: Link Cited by: §I-A. [11] S. Kessler, B. Thomas, and S. Karout (2022) An adapter based pre-training for efficient and scalable self-supervised speech representation learning. In 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 3179–3183. External Links: Document Cited by: §I-C3. [12] M. Li, Y. Ahmadiadli, and X.-P. Zhang (2025-02) A survey on speech deepfake detection. ACM Comput. Surv. 57 (7). External Links: ISSN 0360-0300, Link, Document Cited by: §I. [13] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie (2022-06) A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 11976–11986. Cited by: §IV-A, TABLE I. [14] H. Ma, J. Yi, J. Tao, Y. Bai, Z. Tian, and C. Wang (2021) Continual learning for fake audio detection. In Interspeech 2021, Brno, Czech Republic, p. 886–890. External Links: Document, ISSN 2958-1796 Cited by: §I, §I. [15] D. Mari, F. Latora, and S. Milani (2022) The sound of silence: efficiency of first digit features in synthetic audio detection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), Vol. , Shanghai, China, p. 1–6. External Links: Document Cited by: §I-A. [16] D. Mari, D. Salvi, P. Bestagini, and S. Milani (2025) All-for-one and one-for-all: deep learning-based feature fusion for synthetic speech detection. In Machine Learning and Principles and Practice of Knowledge Discovery in Databases, R. Meo and F. Silvestri (Eds.), Cham, p. 469–479. External Links: ISBN 978-3-031-74627-7 Cited by: §I-A. [17] M. Mulimani and A. Mesaros (2024) Class-incremental learning for multi-label audio classification. In 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 916–920. External Links: Document Cited by: §I. [18] N. Müller, P. Czempin, F. Diekmann, A. Froghyar, and K. Böttinger (2022) Does audio deepfake detection generalize?. In Interspeech 2022, p. 2783–2787. External Links: Document, ISSN 2958-1796 Cited by: 3rd item. [19] V. Negroni, D. Salvi, A. I. Mezza, P. Bestagini, and S. Tubaro (2025) Leveraging mixture of experts for improved speech deepfake detection. In 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 1–5. External Links: Document Cited by: §I, §I. [20] Q. Pham, C. Liu, and S. Hoi (2022) Continual normalization: rethinking batch normalization for online continual learning. External Links: 2203.16102, Link Cited by: §I-C2. [21] R. Reimao and V. Tzerpos (2019) FoR: a dataset for synthetic speech detection. In 2019 International Conference on Speech Technology and Human-Computer Dialogue (SpeD), Vol. , p. 1–10. External Links: Document Cited by: 2nd item. [22] O. Ronneberger, P. Fischer, and T. Brox (2015) U-net: convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi (Eds.), Cham, p. 234–241. External Links: ISBN 978-3-319-24574-4 Cited by: §I-A. [23] D. Salvi, V. Negroni, L. Bondi, P. Bestagini, and S. Tubaro (2025) Freeze and learn: continual learning with selective freezing for speech deepfake detection. In 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 1–5. External Links: Document Cited by: §I, §I-C1, §IV-B, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, Acknowledgment. [24] B. Sun and K. Saenko (2016) Deep coral: correlation alignment for deep domain adaptation. In Computer Vision – ECCV 2016 Workshops, G. Hua and H. Jégou (Eds.), Cham, p. 443–450. External Links: ISBN 978-3-319-49409-8 Cited by: §I-C3. [25] M. Tan and Q. Le (2019-09–15 Jun) EfficientNet: rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, K. Chaudhuri and R. Salakhutdinov (Eds.), Proceedings of Machine Learning Research, Vol. 97, p. 6105–6114. External Links: Link Cited by: §IV-A, TABLE I, TABLE I. [26] M. Todisco, X. Wang, V. Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T.H. Kinnunen, and K. A. Lee (2019) ASVspoof 2019: future horizons in spoofed and fake audio detection. In Interspeech 2019, p. 1008–1012. External Links: Document, ISSN 2958-1796 Cited by: 1st item. [27] C. Wang, J. Yi, X. Zhang, J. Tao, L. Xu, and R. Fu (2023) Low-rank adaptation method for wav2vec2-based fake audio detection. In Proceedings of IJCAI 2023 Workshop on Deepfake Audio Detection and Analysis, External Links: Link Cited by: §I. [28] J. Yi, R. Fu, J. Tao, S. Nie, H. Ma, C. Wang, T. Wang, Z. Tian, Y. Bai, C. Fan, et al. (2022) ADD 2022: the first audio deep synthesis detection challenge. In 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 9216–9220. External Links: Document Cited by: 4th item. [29] X. Zhang, J. Yi, C. Wang, C. Y. Zhang, S. Zeng, and J. Tao (2024-03) What to remember: self-adaptive continual learning for audio deepfake detection. Proceedings of the AAAI Conference on Artificial Intelligence 38 (17), p. 19569–19577. External Links: Link, Document Cited by: §I. [30] Y. Zhang, Z. Zhang, M. Liao, S. Tian, W. Zou, L. Zhang, and C. Xu (2025) Prototypical progressive alignment and reweighting for generalizable semantic segmentation. IEEE Transactions on Intelligent Transportation Systems (), p. 1–17. External Links: Document Cited by: §I-C3.