Personalized Federated Learning Privacy with Adaptive Distillation and Robust Aggregation
Abstract
Federated learning has been shown to alleviate the “data silo” problem; however, it is still vulnerable to threats such as gradient leakage and Byzantine attacks. To address these concerns, a personalized federation privacy protection scheme based on adaptive distillation and robust aggregation has been proposed. This scheme involves the construction of a learnable temperature parameter training strategy in the distillation process, which involves the participation of the local model in the gradient update of the global model. This approach enables the personalization of the local model while ensuring the privacy of the local client. Second, a short-term history gradient is introduced in the aggregation phase to filter low-quality clients, and a paradigm-based adaptive gradient filtering dynamically adjusts the threshold to filter malicious gradients to improve the global model accuracy. Finally, the security analysis shows that honest and curious servers cannot recover data from gradients, and the aggregation phase effectively reduces the impact of unreliable and malicious clients on the global model. The experimental results demonstrate that the proposed scheme significantly enhances the global model accuracy under various types of attacks, including both typical and complex attacks, under both IID and Non-IID conditions.
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.