FLGuard: Robust Federated Learning Using Dual Attention and Gradient Reconstruction

We propose FLGuard, a robust aggregation method for federated learning that enhances resilience to Byzantine attacks. FLGuard integrates cosine sim-ilarity and gradient reconstruction errors computed using a Variational Autoencoder (VAE) within a dual attention mechanism, leveraging a trusted root dataset at the server. By assigning adaptive trust scores to client updates based on these metrics, FLGuard im-proves robustness against advanced attack strategies. Experiments on the MNIST dataset show that FLGuard outperforms existing methods, including FLTrust, across various attack scenarios, achieving faster convergence and higher accuracy while effectively mitigating the impact of malicious clients.