Answers are grounded in this paper and cite their sources.
Personalized federated distillation learning (PFDL) mitigates data and system heterogeneity in edge computing. However, existing PFDL frameworks face communication overhead and privacy leakage risks during sample-level logit exchange. Furthermore, introducing explicit differential privacy noise often substantially degrades model utility. To address these challenges, we propose GDPFDL, a personalized federated distillation framework that utilizes generative networks and a dynamic coordination mechanism. GDPFDL defines the knowledge carrier as a compressed O ( C 2 ) class prototype matrix. It employs a contrastive conditional generative adversarial network (CCGAN) to capture class-level statistical manifolds, thereby mitigating the risk of sample memorization by the generator. Additionally, GDPFDL introduces a dynamic coordination mechanism that adaptively regulates the knowledge upload frequency of each node by quantifying the performance difference between the local model and the global baseline. This mechanism enhances convergence stability in heterogeneous environments and reduces system communication overhead. We evaluate GDPFDL on a simulation platform and a heterogeneous edge testbed comprising real mobile devices using the MNIST, CIFAR, and UCI-HAR datasets. Experimental results indicate that GDPFDL improves accuracy by 1.89% over the optimal baseline on CIFAR-100 under severe non-IID conditions. The framework restricts membership inference attack accuracy to approximately 52.4% and effectively mitigates data reconstruction attacks while maintaining a prototype generation latency per communication round below 100 ms. These findings suggest that GDPFDL operates efficiently in resource-constrained edge environments while preserving user privacy.
No takes yet. Share an insight, caveat, or question.
Duan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: