FedSCAM (Federated Sharpness-Aware Minimization with Clustered Aggregation and Modulation): Scam-resistant SAM for Robust Federated Optimization in Heterogeneous Environments

TL;DR

FedSCAM addresses non-IID heterogeneity in federated learning by per-client Sharpness-Aware Minimization (SAM) radius modulation and heterogeneity-aware aggregation. It estimates a client heterogeneity score from early gradients, scales each client’s SAM radius inversely to this score, and weights updates by both heterogeneity and alignment with the global trajectory, with optional clustering to dampen conflicts. Empirical results on CIFAR-10 and Fashion-MNIST under Dirichlet label skew show FedSCAM achieving competitive final accuracy and improved stability, often with lower compute than SAM-heavy baselines. The work highlights that adaptive, client-specific regularization combined with reliability-aware aggregation offers a practical, scalable path to robust FL in heterogeneous environments.

Abstract

Federated Learning (FL) enables collaborative model training across decentralized edge devices while preserving data privacy. However, statistical heterogeneity among clients, often manifested as non-IID label distributions, poses significant challenges to convergence and generalization. While Sharpness-Aware Minimization (SAM) has been introduced to FL to seek flatter, more robust minima, existing approaches typically apply a uniform perturbation radius across all clients, ignoring client-specific heterogeneity. In this work, we propose \textbf{FedSCAM} (Federated Sharpness-Aware Minimization with Clustered Aggregation and Modulation), a novel algorithm that dynamically adjusts the SAM perturbation radius and aggregation weights based on client-specific heterogeneity scores. By calculating a heterogeneity metric for each client and modulating the perturbation radius inversely to this score, FedSCAM prevents clients with high variance from destabilizing the global model. Furthermore, we introduce a heterogeneity-aware weighted aggregation mechanism that prioritizes updates from clients that align with the global optimization direction. Extensive experiments on CIFAR-10 and Fashion-MNIST under various degrees of Dirichlet-based label skew demonstrate that FedSCAM achieves competitive performance among state-of-the-art baselines, including FedSAM, FedLESAM, etc. in terms of convergence speed and final test accuracy.

Figures (27)

• Figure 1: Mean $\rho$ vs. communication rounds for FedSCAM on Fashion-MNIST, with $\rho _{max}$ set to 0.05
• Figure 2: SAM-family methods on CIFAR-10 under extreme heterogeneity ($\alpha=0.1$).
• Figure 3: SAM variants on CIFAR-10 ($\alpha=0.5$).
• Figure 4: Wall-clock time per communication round on Fashion-MNIST. FedSAM is faster than FedLESAM.
• Figure 5: Aggregation-focused baselines on CIFAR-10 ($\alpha=0.1$).