<p>To address the performance degradation and increased communication overhead that often arise in federated learning under dual heterogeneity of data distributions and model architectures, we propose FedPAAD, a hybrid federated learning framework based on prototype alignment and adaptive distillation. Methodologically: (1) we introduce a reversible dynamic projection for heterogeneous backbones, coupled with Hessian spectrum–guided dimensional adaptation, to map local features into a shared space for robust class-prototype computation and cross-client alignment; (2) on the head network, we employ bidirectional knowledge distillation–using global knowledge to mitigate distribution shift while preserving personalization via local “dark knowledge”; and (3) we design a two-stage weight scheduling mechanism further modulated by distribution divergence, emphasizing prototype learning in early rounds and distillation in later rounds.Theoretically, under a non-convex setting with bounded variance and explicitly accounting for errors from projection and low-rank compression, we derive an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(o\left( \frac{1}{T}\right) \)</EquationSource> </InlineEquation> upper bound on the convergence of the average gradient norm. Empirically, across CIFAR-10/100, Flowers102, and Tiny-ImageNet with both data and system heterogeneity, FedPAAD achieves higher accuracy and greater robustness at low participation rates than strong baselines, while significantly reducing communication at matched accuracy. Overall, FedPAAD attains a superior trade-off among communication efficiency, generalization, and personalization.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Hybrid federated learning with prototype alignment and adaptive distillation

  • Miao Liu,
  • Keming Huang,
  • Yiyang Chen,
  • Zhenxing Sun

摘要

To address the performance degradation and increased communication overhead that often arise in federated learning under dual heterogeneity of data distributions and model architectures, we propose FedPAAD, a hybrid federated learning framework based on prototype alignment and adaptive distillation. Methodologically: (1) we introduce a reversible dynamic projection for heterogeneous backbones, coupled with Hessian spectrum–guided dimensional adaptation, to map local features into a shared space for robust class-prototype computation and cross-client alignment; (2) on the head network, we employ bidirectional knowledge distillation–using global knowledge to mitigate distribution shift while preserving personalization via local “dark knowledge”; and (3) we design a two-stage weight scheduling mechanism further modulated by distribution divergence, emphasizing prototype learning in early rounds and distillation in later rounds.Theoretically, under a non-convex setting with bounded variance and explicitly accounting for errors from projection and low-rank compression, we derive an \(o\left( \frac{1}{T}\right) \) upper bound on the convergence of the average gradient norm. Empirically, across CIFAR-10/100, Flowers102, and Tiny-ImageNet with both data and system heterogeneity, FedPAAD achieves higher accuracy and greater robustness at low participation rates than strong baselines, while significantly reducing communication at matched accuracy. Overall, FedPAAD attains a superior trade-off among communication efficiency, generalization, and personalization.