LiBRA: A Secure Lite-weight Byzantine-Resilient Aggregation Framework

Federated Learning (FL) enables the decentralized training of machine learning models while preserving data privacy. However, this same privacy-preserving paradigm complicates efforts to mitigate malicious or Byzantine clients seeking to disrupt training or degrade model performance. Existing Byzantine-robust aggregation methods often rely on direct inspection of model parameters, leading to privacy risks, computational inefficiencies, and limited adaptability to heterogeneous systems.We propose LiBRA (Lightweight ByzantineRobust Aggregation), a novel aggregation algorithm that satisfies the (α,f)-Byzantine resiliency condition without relying on direct parameter inspection. Instead, LiBRA leverages client-model logits, generated in response to optimized pseudo-pattern inputs, for similarity scoring. This paradigm shift eliminates privacy risks associated with parameter access while significantly reducing computational overhead. LiBRA achieves an asymptotic complexity of O(n2 · (C + logn)), replacing the dominant parameter dimension d (typically millions or billions) with the much smaller model-output dimension C (often in the tens or hundreds).In addition to privacy preservation and computational efficiency, LiBRA seamlessly adapts to heterogeneous systems by leveraging logits for scoring. We extend this approach to introduce the krum-lite, mkrum-lite and bulyan-lite variants, which maintain Byzantine resilience. Through comprehensive experiments, we demonstrate that LiBRA effectively mitigates adversarial attacks while advancing FL toward a more privacy-robust and efficient future.