DPFL-AES: Differential Privacy Federated Learning Based on Adam Early Stopping

Federated learning (FL) is a distributed machine learning framework that protects data privacy on the premise that the data does not leave the local area, and can effectively solve the data islands and data privacy security issues faced by artificial intelligence. However, attackers can launch attribute inference attacks or membership inference attacks on customers by analyzing the parameters in shared model training. To solve this problem, we combine the concepts of differential privacy (DP) and federated learning, and propose a differentially private federated learning scheme based on the Adam Early Stopping algorithm (DPFL-AES). The algorithm inherits the advantages of differential privacy and federated learning, and hides the basic parameters uploaded by the client during the training process. Furthermore, as the model converges, the noise should be reduced accordingly later in the federated learning training process. We added an early stopping (ES) mechanism to the algorithm. In the later stage of federated learning training, as the model converges, the number of communication rounds and the addition of noise are reduced, thereby improving the performance of the model. The algorithm balances data security and availability under a shared model. Finally, we verified the feasibility and practicability of the algorithm through theoretical analysis and experimental simulation on the real MNIST dataset.