Adaptive Gradient Clipping for Robust Federated Learning
This work addresses robustness issues in federated learning for distributed systems, offering an incremental improvement over existing static clipping methods.
The paper tackles the problem of inconsistent robustness in federated learning against adversarial attacks by proposing an adaptive gradient clipping strategy, which significantly enhances robustness in heterogeneous and adversarial settings as confirmed by experiments on benchmark image classification tasks.
Robust federated learning aims to maintain reliable performance despite the presence of adversarial or misbehaving workers. While state-of-the-art (SOTA) robust distributed gradient descent (Robust-DGD) methods were proven theoretically optimal, their empirical success has often relied on pre-aggregation gradient clipping. However, existing static clipping strategies yield inconsistent results: enhancing robustness against some attacks while being ineffective or even detrimental against others. To address this limitation, we propose a principled adaptive clipping strategy, Adaptive Robust Clipping (ARC), which dynamically adjusts clipping thresholds based on the input gradients. We prove that ARC not only preserves the theoretical robustness guarantees of SOTA Robust-DGD methods but also provably improves asymptotic convergence when the model is well-initialized. Extensive experiments on benchmark image classification tasks confirm these theoretical insights, demonstrating that ARC significantly enhances robustness, particularly in highly heterogeneous and adversarial settings.