Paper deep dive
Collaborative Adaptive Curriculum for Progressive Knowledge Distillation
Jing Liu, Zhenchao Ma, Han Yu, Bobo Ju, Wenliang Yang, Chengfang Li, Bo Hu, Liang Song
Intelligence
Status: succeeded | Model: anthropic/claude-sonnet-4.6 | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/24/2026, 2:07:44 AM
Summary
This paper proposes Federated Adaptive Progressive Distillation (FAPD), a consensus-driven framework for collaborative knowledge distillation in federated learning. FAPD uses PCA-based hierarchical decomposition of teacher features, a consensus-driven curriculum controller that monitors global accuracy stability, and client-side progressive distillation with adaptive projection matrices. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet show FAPD achieves 3.64% accuracy improvement over FedAvg on CIFAR-10, 2x faster convergence, and robust performance under extreme data heterogeneity.
Entities (36)
Relation Signals (28)
FAPD → evaluatedon → Tiny-ImageNet
confidence 99% · 45.35% on Tiny-ImageNet
FAPD → evaluatedon → CIFAR-10
confidence 99% · On CIFAR-10, it achieves 89.42% accuracy
FAPD → evaluatedon → CIFAR-100
confidence 99% · FAPD attaining 63.84% on CIFAR-100
FAPD → outperforms → FedAvg
confidence 99% · it attains 3.64% accuracy improvement over FedAvg on CIFAR-10
FAPD → uses → PCA-based Hierarchical Knowledge Decomposition
confidence 99% · FAPD hierarchically decomposes teacher features via PCA-based structuring, extracting principal components ordered by variance contribution
FAPD → uses → Consensus-Driven Curriculum Controller
confidence 99% · a Consensus-Driven Curriculum (CDC) controller dynamically adjusts the knowledge complexity kt for each round t
FAPD → uses → Progressive Knowledge Distillation
confidence 99% · during client-side Progressive Knowledge Distillation (PKD), each client projects its local features
FAPD → addresses → Federated Learning
confidence 98% · FAPD, a novel framework that orchestrates collaborative knowledge transfer through a consensus-driven curriculum
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent advances in collaborative knowledge distillation have demonstrated cutting-edge performance for resource-constrained distributed multimedia learning scenarios. However, achieving such competitiveness requires addressing a fundamental mismatch: high-dimensional teacher knowledge complexity versus heterogeneous client learning capacities, which currently prohibits deployment in edge-based visual analytics systems. Drawing inspiration from curriculum learning principles, we introduce Federated Adaptive Progressive Distillation (FAPD), a consensus-driven framework that orchestrates adaptive knowledge transfer. FAPD hierarchically decomposes teacher features via PCA-based structuring, extracting principal components ordered by variance contribution to establish a natural visual knowledge hierarchy. Clients progressively receive knowledge of increasing complexity through dimension-adaptive projection matrices. Meanwhile, the server monitors network-wide learning stability by tracking global accuracy fluctuations across a temporal consensus window, advancing curriculum dimensionality only when collective consensus emerges. Consequently, FAPD provably adapts knowledge transfer pace while achieving superior convergence over fixed-complexity approaches. Extensive experiments on three datasets validate FAPD's effectiveness: it attains 3.64% accuracy improvement over FedAvg on CIFAR-10, demonstrates 2x faster convergence, and maintains robust performance under extreme data heterogeneity ({\alpha}=0.1), outperforming baselines by over 4.5%.
Tags
Links
- Source: https://arxiv.org/abs/2603.20296v1
- Canonical: https://arxiv.org/abs/2603.20296v1
Trouble viewing inline? Open PDF directly →
Full Text
35,820 characters extracted from source content.
Expand or collapse full text
Collaborative Adaptive Curriculum for Progressive Knowledge Distillation Jing Liu1,2,3,† , Zhenchao Ma2,† , Han Yu1,† , Bobo Ju5, Wenliang Yang1, Chengfang Li4,*, Bo Hu1,*, Liang Song1,* † contribution. *Corresponding authors. This work is supported in part by the National Key Research and Development Program of China under Project No. 2024YFE0200700 (Subject No. 2024YFE0200703). This work was also supported in part by the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, in part by the Shanghai Key Research Laboratory of NSAI and the Joint Laboratory on Networked AI Edge Computing, Fudan University-Changan. Abstract Recent advances in collaborative knowledge distillation have demonstrated cutting-edge performance for resource-constrained distributed multimedia learning scenarios. However, achieving such competitiveness requires addressing a fundamental mismatch: high-dimensional teacher knowledge complexity versus heterogeneous client learning capacities, which currently prohibits deployment in edge-based visual analytics systems. Drawing inspiration from curriculum learning principles, we introduce Federated Adaptive Progressive Distillation (FAPD), a consensus-driven framework that orchestrates adaptive knowledge transfer. FAPD hierarchically decomposes teacher features via PCA-based structuring, extracting principal components ordered by variance contribution to establish a natural visual knowledge hierarchy. Clients progressively receive knowledge of increasing complexity through dimension-adaptive projection matrices. Meanwhile, the server monitors network-wide learning stability by tracking global accuracy fluctuations across a temporal consensus window, advancing curriculum dimensionality only when collective consensus emerges. Consequently, FAPD provably adapts knowledge transfer pace while achieving superior convergence over fixed-complexity approaches. Extensive experiments on three datasets validate FAPD’s effectiveness: it attains 3.64% accuracy improvement over FedAvg on CIFAR-10, demonstrates 2× faster convergence, and maintains robust performance under extreme data heterogeneity (α=0.1), outperforming baselines by over 4.5%. I Introduction Federated Learning (FL) enables privacy-preserving visual model training on decentralized multimedia data [1], critical for visual recognition applications on resource-constrained edge devices in domains like mobile image classification [2], healthcare [3, 4], and video surveillance [5]. A core challenge in FL is managing statistical heterogeneity [6] and communication overhead while accommodating limited client resources [7, 8]. To address these issues, Collaborative Knowledge Distillation (CKD) allows smaller client models to learn from a powerful, centralized teacher model [9], thereby transferring complex knowledge without sharing raw data while striking a balance between model performance and the practical constraints of decentralized systems. Despite its potential, CKD often fails due to a fundamental mismatch between the complexity of teacher-provided knowledge and the learning capacity of heterogeneous clients [1]. Many existing methods [10] attempt to transfer a full, high-dimensional teacher representation from the start of training, yet such “one-size-fits-all” approaches can overwhelm clients with limited resources, leading to unstable training and poor generalization [11, 12]. Other approaches use rigid, predefined curricula, but these static schedules cannot adapt to the dynamic learning states of clients or the collective network [13, 14], proving ineffective in diverse federated settings. The critical gap in current CKD methods is the absence of an adaptive mechanism to orchestrate knowledge transfer based on the network’s collective learning progress. Data-free distillation approaches [15] synthesize training samples but employ fixed knowledge complexity. Other methods [16] distill aggregated statistics without progressive adaptation, while [17] aggregates hyper-knowledge at constant dimensionality. An intelligent system must pace the introduction of complex information, ensuring foundational knowledge is mastered before more intricate details are introduced [2, 18]. Accomplishing this requires a mechanism to monitor the network-wide learning state (i.e., consensus) and use this signal to dynamically adjust the curriculum. Without such adaptability, federated systems cannot efficiently transfer knowledge or accommodate client heterogeneity, leaving performance potential untapped. Existing federated learning can be broadly categorized based on its knowledge transfer strategies [19, 20]. Traditional approaches like FedAvg [21] focus on parameter aggregation without leveraging external teacher knowledge, often struggling with statistical heterogeneity. To address this, FKD methods [9] introduce teacher supervision but typically transfer full, high-dimensional representations throughout training, which can overwhelm resource-constrained clients. More recently, curriculum-based strategies [22] have attempted to structure learning by scheduling data or client participation. However, these methods predominantly rely on linear, predetermined schedules that fail to adapt to the dynamic, collective learning states of the network [23, 24], leaving a critical gap in orchestrating knowledge complexity according to real-time system capacity. Recent advancements have explored integrating large-scale generative models and curriculum strategies into federated learning. Works such as FedDifRC [25] and VQ-FedDiff [26] demonstrate techniques for adapting text-to-image diffusion models to federated settings, addressing communication efficiency. Meanwhile, curriculum learning approaches have been applied to schedule data samples [27] or client participation [22, 28]. Other efforts focus on parameter-efficient adaptation using prompt tuning [29] or collaborative pre-training [11]. While these studies highlight the potential of sophisticated teacher models and curriculum strategies, they do not resolve how to progressively distill complex high-dimensional feature representations to resource-constrained clients [30]. To bridge this gap, we propose Federated Adaptive Progressive Distillation (FAPD), a novel framework that orchestrates collaborative knowledge transfer through a consensus-driven curriculum. Unlike static approaches that overwhelm clients with complex information or rigid schedules that ignore learning dynamics, FAPD dynamically adapts the complexity of distilled knowledge to the network’s collective capacity. Our approach introduces a hierarchical decomposition strategy that structures teacher features into ordered levels of importance, prioritizing fundamental patterns before intricate details in a manner analogous to human educational curricula [31]. A central controller monitors network-wide learning stability and advances the curriculum only when consensus readiness is achieved, thereby ensuring that clients in heterogeneous environments are not destabilized by premature complexity. By synchronizing knowledge transfer with collective progress, FAPD effectively balances the trade-off between learning stability and performance, offering a robust solution for real-world federated systems where client resources and data distributions vary significantly. Our contributions are summarized as follows: • We propose FAPD, a consensus-driven curriculum framework that dynamically orchestrates knowledge complexity in collaborative distillation by monitoring network-wide stability signals and adapting transfer pace accordingly. • We design a PCA-based hierarchical knowledge decomposition that structures teacher features into variance-ordered principal components, enabling progressive distillation synchronized with heterogeneous client learning. • We conduct extensive experiments that demonstrate FAPD achieves substantial improvements over baselines in accuracy, convergence speed, and training stability under diverse heterogeneous federated learning environments. Figure 1: Illustration of FAPD featuring three components: (a) server-side hierarchical knowledge decomposition via PCA generating rotation matrix R, (b) consensus-driven adaptive curriculum controller monitoring accuracy history A to determine dimension ktk_t, and (c) client-side progressive knowledge distillation projecting features through PtP_t to compute multi-objective loss ℒcL_c. I Related Work Collaborative Knowledge Distillation. CKD enables resource-constrained clients to learn from server models while preserving data privacy. Early methods required public datasets [9], while data-free alternatives synthesize samples via generative models [15] or distill aggregated statistics [17]. Recent frameworks address architectural heterogeneity through group distillation [10] or robust transfer protocols [3]. However, existing approaches uniformly employ fixed-complexity transfer strategies, ignoring dynamic client learning capacity. In contrast, FAPD introduces an adaptive curriculum over distilled knowledge complexity. Curriculum Learning in Federated Settings. Curriculum Learning improves convergence by training on progressively difficult examples [27]. Within FL, prevailing methods schedule data samples by difficulty metrics [32], manage client participation [22], or guide personalization from generic to client-specific features [28]. Nevertheless, such methods schedule data or clients rather than knowledge complexity. FAPD introduces curriculum over distilled knowledge dimensionality, paced by network-wide consensus signals reflecting collective learning states. I Preliminary Federated Knowledge Distillation. In a FL system with C clients, each client c∈1,…,Cc∈\1,…,C\ holds a private dataset c=(i,yi)i=1ncD_c=\(x_i,y_i)\_i=1^n_c. The global objective is to minimize a weighted average of local loss without centralizing data: minwF(w):=∑c=1CpcFc(w), _wF(w):= _c=1^Cp_cF_c(w), (1) where pc=nc/∑jnjp_c=n_c/ _jn_j is the weight for client c, and Fc(w)=1nc∑(,y)∈cℒ(f(;w),y)F_c(w)= 1n_c _(x,y) _cL(f(x;w),y) is the local empirical risk. In FKD, a powerful teacher model with parameters wTw_T assists the training of smaller client (student) models with parameters wS,cw_S,c. Knowledge is transferred by aligning feature representations. Let gT(⋅;wT)g_T(·;w_T) and gS(⋅;wS,c)g_S(·;w_S,c) be the feature extractors of the teacher and student models, respectively. The local objective for each client c is a combination of a standard classification loss and a distillation loss: minwS,c(,y)∼c[ℒCE(fS(;wS,c),y)+λℒKD], _w_S,cE_(x,y) _c [L_CE(f_S(x;w_S,c),y)+ _KD ], (2) where fSf_S is the student’s prediction function, ℒCEL_CE is the cross-entropy loss, and ℒKD=Dist(gS(;wS,c),gT(;wT))L_KD=Dist(g_S(x;w_S,c),g_T(x;w_T)) is a loss function, such as Mean Squared Error, that aligns student and teacher features. Hierarchical Knowledge Representation via PCA. A primary challenge in FKD is that high-dimensional teacher representations, T=gT(;wT)∈ℝDz_T=g_T(x;w_T) ^D, can overwhelm student models. To address this, knowledge can be structured and introduced progressively. PCA provides a principled method for decomposing a feature space into a hierarchy of components ordered by their contribution to the data’s variance [33, 34]. Given a set of teacher features, PCA finds an orthogonal transformation to a new coordinate system of principal components, where projecting a feature vector Tz_T onto the first k principal components provides a compressed representation that captures the most significant variations in the data. By projecting Tz_T onto the subspace spanned by the eigenvectors 1,…,k\v_1,…,v_k\ of the data’s covariance matrix (corresponding to the k largest eigenvalues), we obtain the k-dimensional representation: T(k)=[1,…,k]TT.z_T^(k)=[v_1,…,v_k]^Tz_T. (3) This provides a principled foundation for a curriculum that introduces knowledge of increasing complexity, forming the core of our adaptive distillation strategy. IV Method IV-A Overview of FAPD The FAPD framework, illustrated in Fig. 1, orchestrates adaptive knowledge transfer through three stages. First, the server performs Hierarchical Knowledge Decomposition (HKD) on the teacher’s high-dimensional features T∈ℝDz_T ^D using PCA, generating an orthogonal rotation matrix R that orders dimensions by variance. Second, a Consensus-Driven Curriculum (CDC) controller dynamically adjusts the knowledge complexity ktk_t for each round t. By monitoring the global accuracy history A, the controller evaluates a stability condition C(t)C(t) to determine if the network has reached a consensus, advancing ktk_t when stability is confirmed. Finally, during client-side Progressive Knowledge Distillation (PKD), each client projects its local features S,cz_S,c and the teacher’s features into the current ktk_t-dimensional subspace using the projection matrix PtP_t. Clients then optimize a composite objective ℒcL_c that aligns these projected representations, ensuring that knowledge transfer is synchronized with the network’s collective learning capacity. IV-B Hierarchical Knowledge Decomposition To structure the knowledge transfer, we decompose the teacher’s D-dimensional feature space into an ordered hierarchy. We apply PCA to a representative set of teacher feature embeddings, T,i∈ℝDi=1M\z_T,i ^D\_i=1^M, extracted from a calibration dataset, where M denotes the number of calibration samples. The data covariance matrix is: Σ=1M−1∑i=1M(T,i−¯T)(T,i−¯T)T, = 1M-1 _i=1^M(z_T,i- z_T)(z_T,i- z_T)^T, (4) where ¯T z_T is the mean feature vector. Eigendecomposition of Σ yields a set of orthogonal principal components: Σ=VΛVT, =V V^T, (5) where V=[1,…,D]V=[v_1,…,v_D] is the matrix of eigenvectors (principal components) and Λ is a diagonal matrix of eigenvalues λ1,…,λD\ _1,…, _D\, sorted such that λ1≥⋯≥λD _1≥…≥ _D. Consequently, the matrix R=VT∈ℝD×DR=V^T ^D× D serves as a global rotation matrix. At curriculum stage t, the projection matrix for dimensionality ktk_t is constructed by selecting the first ktk_t rows of R: Pt=R[:kt,:]∈ℝkt×D.P_t=R[:k_t,:] ^k_t× D. (6) Consequently, this matrix projects any feature vector onto the ktk_t-dimensional subspace that captures the most data variance. IV-C Consensus-Driven Curriculum Controller The central innovation of FAPD is the server-side curriculum controller that dynamically paces knowledge transfer, where the controller adjusts the feature dimension ktk_t based on a network-wide stability consensus. At the end of each round t, the server evaluates the global accuracy, acctacc_t, and the network is considered to have reached consensus if the accuracy has plateaued. Formally, let ϵε be a stability threshold and N be the consensus window size; the stability condition C(t)C(t) is met if all recent accuracies are close to the current accuracy: C(t)=⋀τ=t−N+1t−1(|acct−accτ|<ϵ).C(t)= _τ=t-N+1^t-1(|acc_t-acc_τ|<ε). (7) The curriculum dimension for the next round, kt+1k_t+1, is then updated according to the rule: kt+1=min(kt+Δk,D)if C(t) is truektotherwise,k_t+1= cases (k_t+ k,D)&if C(t) is true\\ k_t&otherwise cases, (8) where k0k_0 is the initial dimension and Δk k is the step size. By implementing this curriculum control, more complex knowledge is introduced only when the network has collectively mastered the current representations. IV-D Client-Side Progressive Distillation During local training at round t, each client c receives the global model parameters wtw_t and curriculum dimension ktk_t. For each input x, the client computes its feature representation S,c=gS(;wS,c)z_S,c=g_S(x;w_S,c). To align with the teacher, features are projected into the ktk_t-dimensional subspace using PtP_t from Eq. 6, yielding S,ckt=PtS,cz_S,c^k_t=P_tz_S,c and Tkt=PtTz_T^k_t=P_tz_T, where Tz_T denotes the teacher’s feature representation for the same input. The local training objective for client c combines three components: ℒc=ℒCE+λkdℒKD+λclℒCL.L_c=L_CE+ _kdL_KD+ _clL_CL. (9) Here, the classification loss ℒCEL_CE is the standard cross-entropy on ground-truth labels, while the knowledge distillation loss ℒKDL_KD enforces consistency between student and teacher feature distributions in the projected space. First, both feature vectors are L2-normalized; then, we compute the KL-divergence: ℒKD=DKL(logsoftmax(^S,ckt)∥softmax(^Tkt)),L_KD=D_KL( ( z_S,c^k_t)\|softmax( z_T^k_t)), (10) where ^=/‖2 z=z/\|z\|_2. The contrastive loss ℒCLL_CL refines the feature space by aligning image features with text-based semantic embeddings, following the InfoNCE framework adapted from recent diffusion-based federated approaches [25]. Let text,yktz_text,y^k_t be the projected text feature for the correct class y (positive) and text,jktj≠y\z_text,j^k_t\_j≠ y be for other classes (negatives). The loss is: ℒCL=−logexp(sim(S,ckt,text,ykt)/τ)∑jexp(sim(S,ckt,text,jkt)/τ),L_CL=- (sim(z_S,c^k_t,z_text,y^k_t)/τ) _j (sim(z_S,c^k_t,z_text,j^k_t)/τ), (11) where τ is a temperature hyperparameter and sim(⋅,⋅)sim(·,·) is the cosine similarity. Through this progressive, multi-faceted objective, clients learn a rich representation aligned with both visual and semantic features. The complete FAPD training procedure is detailed in LABEL:alg:fapd of the supplementary material. V Experiments TABLE I: Accuracy (%) of comparison across datasets (α=0.5). Method CIFAR-10 CIFAR-100 Tiny-ImageNet FedAvg [21] 85.78 61.26 43.35 FedProx [35] 85.68 61.40 43.48 MOON [36] 86.10 61.48 43.78 FedNH [10] 86.25 61.43 44.12 FedRCL [11] 86.89 62.36 44.89 FedCDA [13] 87.11 61.95 44.56 FAPDnadpt 87.23 62.18 44.67 FAPDncont 87.89 63.05 45.28 FAPD (Ours) 89.42 63.84 45.35 V-A Experimental Setup Datasets and Federated Setting. Experiments are conducted on three benchmark datasets: CIFAR-10/100 [37] and Tiny-ImageNet [38], a 200-class subset of ImageNet with 64×6464× 64 images. The federated environment consists of 10 clients, with 5 randomly selected in each of the 100 communication rounds. To simulate statistical heterogeneity, training data is partitioned among clients using a Dirichlet distribution with concentration parameter α∈0.1,0.5,1.0α∈\0.1,0.5,1.0\ controlling non-IID intensity, where smaller values induce more severe distribution skew. Implementation Details. All experiments are conducted using PyTorch on NVIDIA 4090 GPUs. Client models employ ResNet-10 for CIFAR-10 and CIFAR-100, while MobileNetV2 is used for Tiny-ImageNet to ensure fair comparison. The teacher model provides pre-extracted 512-dimensional features from a Stable Diffusion [39] encoder. Each client trains for E∈5,10,20E∈\5,10,20\ local epochs per round using SGD optimizer with learning rate 0.01, momentum 0.9, and batch size 64. FAPD is configured with initial curriculum dimension k0=8k_0=8, step size Δk=5 k=5, stability threshold ϵ=0.005ε=0.005, and consensus window N=3N=3 rounds. The distillation objective combines KL-divergence loss and InfoNCE contrastive loss with temperature τ=0.04τ=0.04, weighted by λkd=0.5 _kd=0.5 and λcl=0.5 _cl=0.5. Baseline Methods. We compare FAPD against eight baselines: FedAvg [21], the standard federated averaging algorithm; FedProx [35], adding proximal terms for heterogeneity handling; MOON [36], employing model-level contrastive learning; FedNH [10], tackling heterogeneity using class prototypes; FedRCL [11], proposing relaxed contrastive learning; FedCDA [13], aggregating cross-round local models; variant model FAPDnadpt, ablation without adaptive mechanism; and variant model FAPDncont, ablation without contrastive learning. Figure 2: Result of local epoch count analysis on CIFAR-10/100. TABLE I: Ablation study on components and loss functions. ID HKD CDC PKD ℒCEL_CE ℒKDL_KD ℒCLL_CL CIFAR-10 CIFAR-100 1 ✗ ✗ ✗ ✓ ✗ ✗ 85.78 61.26 2 ✗ ✗ ✗ ✓ ✗ ✓ 86.23 61.67 3 ✓ ✗ ✗ ✓ ✓ ✗ 86.12 61.74 4 ✓ ✗ ✗ ✓ ✓ ✓ 87.23 62.18 5 ✓ ✓ ✓ ✓ ✓ ✗ 87.89 63.05 6 ✓ ✓ ✓ ✓ ✓ ✓ 89.42 63.84 Figure 3: Result of client scalability analysis on CIFAR-10/100. V-B Results and Analysis Performance Comparison. FAPD demonstrates superior performance across all benchmarks, as shown in Table I. On CIFAR-10, it achieves 89.42% accuracy, outperforming the strongest baseline, FedCDA, by 2.31% and FedAvg by 3.64%. This advantage extends to more complex tasks, with FAPD attaining 63.84% on CIFAR-100 and 45.35% on Tiny-ImageNet, consistently surpassing contrastive methods like FedRCL and MOON. Ablation results highlight the synergy of our approach: removing the adaptive mechanism (FAPDnadpt) or contrastive learning (FAPDncont) leads to significant performance drops of 2.19% and 1.53% on CIFAR-10, respectively. These results confirm that FAPD’s adaptive curriculum effectively manages knowledge complexity, yielding substantial gains over static distillation strategies, particularly as dataset difficulty increases. Scalability Analysis. We analyze the impact of local epochs (E) and client scalability (K) on convergence and accuracy, as illustrated in Fig. 2 and Fig. 3. For local epochs, FAPD achieves optimal performance at E=10E=10 on both CIFAR-10 (89.42%) and CIFAR-100 (63.84%). While increasing E to 20 slightly degrades accuracy (e.g., 88.95% on CIFAR-10) due to potential client drift, FAPD consistently outperforms FedAvg across all E∈5,10,20E∈\5,10,20\, demonstrating robustness to local training variations. Regarding scalability, increasing the total client count K from 10 to 50 introduces greater heterogeneity and data sparsity. Consequently, accuracy naturally declines, yet FAPD maintains a significant lead. On CIFAR-10, FAPD achieves 84.56% with K=50K=50, surpassing FedAvg’s 80.12% by 4.44%. Similarly, on CIFAR-100, FAPD retains 59.78% accuracy at K=50K=50 versus FedAvg’s 55.89%, validating the framework’s effectiveness in larger-scale federated networks. Ablation Study. Table I systematically isolates FAPD’s component contributions. Model 1 (FedAvg/baseline) with only ℒCEL_CE achieves 85.78% on CIFAR-10. Model 2 adds ℒCLL_CL, improving to 86.23%, demonstrating semantic alignment benefits. Model 3 introduces HKD with ℒKDL_KD, reaching 86.12%, while Model 4 (FAPDnadpt) combines HKD with both distillation losses, achieving 87.23%. Model 5 (FAPDncont) integrates all components (HKD, CDC, PKD) but excludes contrastive learning, reaching 87.89% and highlighting consensus-driven adaptive pacing effectiveness. Model 6 (complete FAPD) achieves peak performance of 89.42% on CIFAR-10 and 63.84% on CIFAR-100. Comparing Models 5 and 6 reveals ℒCLL_CL contributes 1.53% gain, while Models 4 and 5 show adaptive curriculum control (CDC+PKD) adds 0.66%. The synergistic combination enables FAPD to effectively balance knowledge complexity with client learning capacity across heterogeneous federated environments. Visualization Analysis. Fig. 4 visualizes the t-SNE embeddings of CIFAR-10 test samples to assess representation quality. FedAvg results in severely entangled clusters with ambiguous boundaries, particularly between similar classes like automobiles and trucks. In contrast, FAPD produces compact, well-separated clusters with distinct inter-class margins. While the non-adaptive variant (FAPDnadpt) shows only slight improvement, the full FAPD framework effectively organizes the feature space, aligning closely with semantic hierarchies. Quantitatively, FAPD achieves 3.64 percentage points higher accuracy than FedAvg on CIFAR-10 (89.42% vs. 85.78%), representing a 4.24% relative improvement, demonstrating that the synergistic combination of hierarchical decomposition and consensus-driven adaptation enables the learning of highly discriminative representations that capture both coarse and fine-grained semantic distinctions. Robustness Analysis. We evaluate FAPD’s resilience to statistical heterogeneity using Dirichlet partitions (α∈0.1,0.2,0.5,1.0α∈\0.1,0.2,0.5,1.0\), as shown in Table I. Under extreme non-IID conditions (α=0.1α=0.1), FAPD maintains 85.87% accuracy on CIFAR-10, outperforming FedAvg (81.35%) by 4.52%. Notably, as heterogeneity increases (α from 1.0 to 0.1), FedAvg suffers a sharp 5.77% drop, whereas FAPD degrades by only 4.31%, demonstrating superior stability. This robustness extends to CIFAR-100, where FAPD achieves 60.12% at α=0.1α=0.1, surpassing FedAvg’s 56.12% by 4.00%. The consistent performance advantage of the full FAPD over variants (FAPDnadpt and FAPDncont) across all α settings confirms that the consensus-driven adaptive curriculum effectively mitigates client drift caused by severe distribution skew. TABLE I: Accuracy (%) under varying α level. Dataset Method α=0.1α=0.1 α=0.2α=0.2 α=0.5α=0.5 α=1.0α=1.0 CIFAR-10 FedAvg [21] 81.35 84.12 85.78 87.12 FAPDnadpt 83.12 85.78 87.23 88.45 FAPDncont 84.56 86.89 87.89 89.12 FAPD (Ours) 85.87 88.34 89.42 90.18 CIFAR-100 FedAvg [21] 56.12 58.45 61.26 62.89 FAPDnadpt 57.45 59.78 62.18 63.56 FAPDncont 58.89 61.23 63.05 64.45 FAPD (Ours) 60.12 62.45 63.84 65.12 Figure 4: t-SNE visualization comparing feature distributions. Discussion. While FAPD demonstrates significant efficacy, several limitations warrant discussion. First, the hierarchical decomposition relies on PCA derived from a calibration dataset. If this data does not adequately represent the global distribution, the resulting knowledge hierarchy may be suboptimal. Second, the current consensus mechanism applies a unified curriculum across the network. In scenarios with extreme resource heterogeneity, a personalized curriculum that adapts to individual client capacities could potentially offer further gains. Finally, our framework is currently optimized for image classification tasks using CNNs. Extending the hierarchical decomposition principle to other multimedia modalities, such as video sequences or audio spectrograms, requires domain-specific adaptations of the variance-based ordering. VI Conclusions We present FAPD, a consensus-driven framework that dynamically orchestrates knowledge transfer complexity in collaborative knowledge distillation. By monitoring network-wide learning stability and progressively expanding feature dimensionality through PCA-based hierarchical decomposition, FAPD adaptively advances curriculum dimensionality when collective consensus emerges, ensuring knowledge complexity matches heterogeneous client capacity. Extensive experiments on three datasets demonstrate FAPD achieves 3.64% accuracy improvement over FedAvg with 2× faster convergence, and exhibits remarkable resilience to data heterogeneity, maintaining high performance even under extreme non-IID settings. Future work will explore client-specific adaptive curricula and multi-modal extensions to video analytics tasks. References [1] K. Pfeiffer, M. Rapp, R. Khalili, and J. Henkel, “Federated learning for computationally-constrained heterogeneous devices: A survey,” ACM Comput. Surv., vol. 55, no. 14s, p. 1–27, 2023. [2] J. Liu, Y. Liu, J. Lin, J. Li, L. Cao, P. Sun, B. Hu, L. Song, A. Boukerche, and V. C. Leung, “Networking systems for video anomaly detection: A tutorial and survey,” ACM Comput. Surv., vol. 57, no. 10, p. 270:1–270:37, 2025. [3] X. Mai, J. Lin, H. Wang, Z. Tao, Y. Wang, S. Yan, X. Tong, J. Yu, B. Wang, Z. Zhou, Q. Zhao, S. Gao, and W. Zhang, “All rivers run into the sea: Unified modality brain-inspired emotional central mechanism,” in ACM M, 2024, p. 632–641. [4] C. Li, G. Fang, Y. Liu, J. Liu, and L. Song, “Decoding silent reading eeg signals using adaptive feature graph convolutional network,” IEEE Signal Process. Lett., vol. 31, p. 1–5, 2024. [5] Y. Liu, J. Liu, K. Yang, B. Ju, S. Liu, Y. Wang, D. Yang, P. Sun, and L. Song, “Amp-net: Appearance-motion prototype network assisted automatic video anomaly detection system,” IEEE Trans. Ind. Informat., vol. 20, no. 2, p. 2843–2855, Feb. 2024. [6] Ehsan Hallaji, R. Razavi-Far, M. Saif, B. Wang, and Q. Yang, “Decentralized federated learning: A survey on security and privacy,” IEEE Trans. Big Data, vol. 10, no. 2, p. 194–213, 2024. [7] Y. Liu, Y. Kang, T. Zou, Y. Pu, Y. He, X. Ye, Y. Ouyang, Y.-Q. Zhang, and Q. Yang, “Vertical federated learning: Concepts, advances, and challenges,” IEEE Trans. Knowl. Data Eng., vol. 36, no. 7, p. 3615–3634, 2024. [8] J. Liu, Y. Du, K. Yang, J. Wu, Y. Wang, X. Hu, Z. Wang, Y. Liu, P. Sun, A. Boukerche, and V. C. Leung, “Edge-cloud collaborative computing on distributed intelligence and model optimization: A survey,” IEEE Commun. Surveys Tuts., p. 1–1, 2026. [9] D. Li and J. Wang, “Fedmd: Heterogenous federated learning via model distillation,” arXiv preprint arXiv:1910.03581, 2019. [10] Y. Dai, Z. Chen, J. Li, S. Heinecke, L. Sun, and R. Xu, “Tackling data heterogeneity in federated learning with class prototypes,” AAAI, vol. 37, no. 6, p. 7314–7322, 2023. [11] S. Seo, J. Kim, G. Kim, and B. Han, “Relaxed contrastive learning for federated learning,” in CVPR, 2024, p. 12 279–12 288. [12] J. Wu, S. Zhang, M. Hou, Z. Wang, W. Chen, Z. Tian, F. R. Yu, and V. C. M. Leung, “Clip-ae: A multi-modal unsupervised images enhancement method based on high-order adaptive curve for visual disbalance defects,” IEEE Trans. Multimedia, vol. 27, p. 4269–4283, 2025. [13] H. Wang, H. Xu, Y. Li, Y. Xu, R. Li, and T. Zhang, “Fedcda: Federated learning with cross-rounds divergence-aware aggregation,” in ICLR, 2023. [14] J. Wu, J. Liu, Y. Liu, L. Wang, Z. Wang, W. Chen, Z. Tian, R. Yu, and V. C. M. Leung, “A survey on cloud-edge-terminal collaborative intelligence in aiot networks,” arXiv.2508.18803, 2025. [15] L. Zhang, L. Shen, L. Ding, D. Tao, and L.-Y. Duan, “Fine-tuning global model via data-free knowledge distillation for non-iid federated learning,” CVPR, p. 10 164–10 173, 2022. [16] S. Zhao, T. Liao, L. Fu, C. Chen, J. Bian, and Z. Zheng, “Data-free knowledge distillation via generator-free data generation for non-iid federated learning,” Neural Netw., vol. 179, p. 106627, 2024. [17] H. Chen, C. Wang, and H. Vikalo, “The best of both worlds: Accurate global and personalized models through federated learning with data-free hyper-knowledge distillation,” ArXiv, vol. abs/2301.08968, 2023. [18] J. Liu, Z. Ma, Z. Wang, Y. Liu, Z. Wang, P. Sun, L. Song, B. Hu, and V. C. M. Leung, “A survey on diffusion models for anomaly detection,” arXiv:2501.11430, 2025. [19] P. Yan and G. Long, “Client-supervised federated learning: Towards one-model-for-all personalization,” in IEEE ICME, 2024, p. 1–6. [20] Y. Wen, Y. Wang, K. Yi, J. Ke, and Y. Shen, “Diffimpute: Tabular data imputation with denoising diffusion probabilistic model,” in IEEE ICME, 2024, p. 1–6. [21] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in AISTATS, 2017, p. 1273–1282. [22] M. Wang, J. Guo, and W. Jia, “Federated multi-phase curriculum learning to synchronously correlate user heterogeneity,” IEEE Trans. Artif. Intell., vol. 5, p. 2026–2039, 2022. [23] Y. Liu, J. Liu, C. Li, R. Xi, W. Li, L. Cao, J. Wang, L. T. Yang, J. Yuan, and W. Zhou, “Anomaly detection and generation with diffusion models: A survey,” arXiv:2506.09368, 2025. [24] J. Liu, G. Fang, L. Teng, L. Qian, B. Hu, and P. Sun, “Enhancing collaborative learning efficiency via control-theoretic merit gating in federated networks,” in IEEE ICC, 2026. [25] H. Wang, H. Li, H. Chen, J. Yan, J. Shi, and J. Shen, “Feddifrc: Unlocking the potential of text-to-image diffusion models in heterogeneous federated learning,” in ICCV, 2025. [26] T. Yoon, M. Hwang, and E. Yang, “Vq-feddiff: Federated learning algorithm of diffusion models with client-specific vector-quantized conditioning,” IEEE Trans. Pattern Anal. Mach. Intell., p. 1–11, 2025. [27] Y. Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” in ICML, 2009, p. 41–48. [28] M. Kang and H. Park, “Curriculum guided personalized subgraph federated learning,” ArXiv, vol. abs/2509.00402, 2025. [29] Y. Bao, “Prompt tuning empowering downstream tasks in multimodal federated learning,” BESC, p. 1–7, 2024. [30] J. Liu, Y. Du, Y. Liu, Z. Wang, P. Sun, and V. C. M. Leung, “Projecting to consensus: Communication-efficient collaborative learning across heterogeneous networks,” in IEEE ICC, 2026. [31] J. Liu, L. Gong, J. Guo, J. Wu, L. Sun, Y. Bi, K. Patwari, B. Chen, L. Zhang, W. Zhou, Y. Liu, X. Zhu, C.-N. Chuah, and B. Rajaratnam, “Multimodal large language models in medicine and nursing: A survey,” techrxiv.175623882.21520632/v1, 2025. [32] Y. Wang, S. Yan, W. Song, A. Liotta, J. Liu, D. Yang, S. Gao, and W. Zhang, “Mgr3 net: Multigranularity region relation representation network for facial expression recognition in affective robots,” IEEE Trans. Ind. Informat., vol. 20, no. 5, p. 7216–7226, May 2024. [33] C. Li, Y. Liu, J. Li, Y. Miao, J. Liu, and L. Song, “Decoding bilingual eeg signals with complex semantics using adaptive graph attention convolutional network,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 32, p. 249–258, 2024. [34] C. Li, H. Wang, Y. Liu, X. Zhu, and L. Song, “Silent eeg classification using cross-fusion adaptive graph convolution network for multilingual neurolinguistic signal decoding,” Biomed. Signal Process. Control, vol. 87, p. 105524, 2024. [35] T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V. Smith, “Federated optimization in heterogeneous networks,” in MLSys, vol. 2, 2020, p. 429–450. [36] Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” in CVPR, 2021, p. 10 713–10 722. [37] A. Krizhevsky, G. Hinton et al., “Convolutional deep belief networks on cifar-10,” vol. 40, no. 7, p. 1–9, 2010. [38] Y. Le and X. Yang, “Tiny imagenet visual recognition challenge,” CS 231N, vol. 7, no. 7, p. 3, 2015. [39] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in CVPR, 2022, p. 10 684–10 695.