Joint Layer Selection and Differential Privacy Design for Federated Learning Over Wireless Networks

In this work, the problem of training the secure federated learning (FL) algorithm over a multicell wireless network is investigated. FL is indeed a learning method that can protect users’ privacy, but it has also been shown to be vulnerable to gradient leakage attacks, which can leak users’ private data. Therefore, we propose a defense method to prevent gradient leakage attacks by uploading the selected layers to attend global model updates. Moreover, we add the differential privacy (DP) noise to the model during the local training to strengthen the defense further. Furthermore, to minimize the total privacy leakage for users, we formulate an optimization problem that minimizes this leakage by jointly optimizing resource block (RB) allocation, layer selection, and DP noise. To find a locally optimal solution to this problem, we divide it into two subproblems: 1) initially solving for RB allocation and 2) layer selection using successive convex approximation (SCA) for convex approximation, followed by optimizing DP noise. The simulation results demonstrate that our proposed optimization algorithm outperforms the other two benchmarks.