Faros: robust federated learning with adaptive scaling against backdoor attacks

Abstract Federated Learning (FL) enables multiple clients to collaboratively train a shared model without exposing local data, making it a fundamental paradigm for large-scale distributed intelligence. However, in practical edge-cloud deployments, the server must inspect a large volume of high-dimensional client updates within tight communication windows, which makes secure aggregation a problem closely tied to parallel processing, real-time response, and high-performance computing (HPC) resources. Among the major threats to FL, backdoor attacks are particularly insidious because they implant malicious behaviors into the global model while preserving benign-task performance. Although pre-aggregation defenses based on gradient analysis are promising, the current state-of-the-art methods such as Scope suffer from two key limitations: fixed parameters are ineffective against adaptive attackers, and single-point clustering is vulnerable to failure under heterogeneous (non-IID) data distributions. To address these limitations, we propose FAROS, a robust and HPC-friendly defense framework that generalizes the transform-and-cluster paradigm. FAROS incorporates two key components: Adaptive Differential Scaling (ADS), which dynamically adjusts defense sensitivity according to the dispersion of client gradients, and Robust Core-set Computing (RCC), which replaces single-point clustering with a consensus-based centroid derived from a stable core-set. This design improves robustness while preserving server-side efficiency through vectorizable similarity computation and parallelizable filtering. Extensive experiments on multiple datasets, models, and attack settings show that FAROS consistently outperforms existing defenses in both attack suppression and benign-task accuracy, while remaining compatible with scalable distributed FL infrastructures.