Adversarial Training for Secure Federated Learning with Differential Privacy

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources without directly exchanging data. However, this paradigm is increasingly vulnerable to adversarial attacks, where malicious participants can manipulate the training process to infer sensitive information about individual users. This paper proposes a novel framework combining adversarial training with differential privacy to mitigate these security risks. Our approach trains models robust against attacks aimed at extracting private data while simultaneously guaranteeing differential privacy for each participant. We introduce a modified training loop incorporating adversarial loss alongside the standard FL loss, and demonstrate its effectiveness through theoretical analysis and a conceptual examination. The key contribution lies in the synergistic combination of these two techniques, creating a more secure and privacy-preserving federated learning system. The goal is to provide a foundational approach for building robust FL systems.