Security Evaluation of Differentially Private Federated Learning Against Model Inversion Attacks

Federated learning has gained increasing attention in social applications in recent years. However, concerns have emerged regarding potential threats to the internal information security of learning models. This research addresses these security challenges through several key contributions: We propose a differential privacy implementation for the Federated Averaging (FedAvg) algorithm. This approach aims to create a secure federated learning environment while preserving the benefits of FedAvg's weighted aggregation. We also investigate attack scenarios involving malicious participants in the federated learning process, with a particular focus on model-reversal attacks, where an adversary attempts to extract learning data from the global model. Finally, we evaluate the effectiveness of security measures designed to counter these model-reversal attacks. Through extensive experimentation, we analyze both the ability to protect differential privacy and the potential vulnerabilities in collaborative learning systems. Our findings provide valuable insights into how to safeguard collaborative learning systems while preserving the utility of real-world applications.