Privacy-Preserving Federated Distillation Resilient to Client Disconnections and Poisoning Attacks
Abstract
Abstract Federated distillation (FD) reduces communication overhead and supports model heterogeneity by sharing prediction knowledge rather than model parameters, but its deployment in edge environments is impeded by non‑IID data, client disconnections, privacy leakage, and poisoning vulnerabilities. This paper proposes GSE‑SFD, a grouped selective FD framework that jointly tackles these challenges via three synergistic components: density‑ratio‑based out‑of‑distribution filtering (KuLSIF) to retain reliable soft labels, threshold secret sharing for grouped secure aggregation resilient to client dropout, and adaptive Euclidean‑distance anomaly detection at the central server to filter tampered group knowledge. Extensive experiments on MNIST, FashionMNIST, and CIFAR‑10 under both label‑skew and quantity‑skew non‑IID settings demonstrate that GSE‑SFD achieves accuracy comparable to Selective‑FD while maintaining over 95% aggregation success under up to 40% client dropout. In addition, it substantially strengthens privacy: reconstructed images from gradient inversion attacks exhibit 2.5× higher MSE and 2.4× lower SSIM than those from Selective‑FD, confirming the effective concealment of individual knowledge. These results, together with the low communication overhead and GAS‑level attack resistance, establish GSE‑SFD as a practical, privacy‑preserving, and dropout‑tolerant solution for resource‑constrained edge learning scenarios, with particular promise for sensitive domains…
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.