APRA-DP: Differential Privacy based Adaptive Privacy Preserving Robust Aggregation Method for Federated Learning

Federated learning is a popular approach in machine learning because it protects participants’ data privacy and leverages collective knowledge to enhance model performance. However, it faces security vulnerabilities and privacy risks, such as feature inference attacks and poisoning attacks. Current methodologies often overlook the interdependence of privacy preservation and security fortification. We propose an adaptive differential privacy robust aggregation methodology that safeguards privacy and provides resilience against poisoning attacks. This involves implementing measures on the client side to mitigate the adverse effects of adding noise through differential privacy to locally trained models before transmission. Additionally, Byzantine robust aggregation strategies are applied on the server side to defend against poisoning attacks. Our proposed approach outperforms prevailing state-of-the-art algorithms through extensive evaluation using various real-world datasets.