ADLG: An Adaptive Deep Leakage from Gradients in Federated Learning
Zizheng Zhang, Lei Mo · 2024
Federated learning has garnered increasing attention due to its privacy-preserving capabilities. However, the unique framework of federated learning results in gradients being transmitted between clients and server, leading to extensive research on attacks that reconstruct private data from gradients information. Nevertheless, existing gradients-based attack methods often need higher accuracy and iteration speed in recovering repeated labels and reconstructing original images. We propose a novel adaptive estimation method to address this issue and accelerate gradient leakage attacks. This algorithm dynamically adjusts parameter estimations of influence quantities based on the number of repeated labels. Compared to existing methods, our algorithm enhances estimation accuracy when encountering excessively repeated labels in a single batch, thereby improving the convergence speed of attacks. Finally, we demonstrate the superiority of our algorithm in terms of convergence speed and accuracy over existing methods through practical federated learning tasks.