Title: GRANITE: A Byzantine-Resilient Dynamic Gossip Learning Framework

Abstract

Gossip Learning (GL) represents a decentralized machine learning approach in which participants periodically share and combine model updates with a limited number of adjacent peers. While recent methodologies have leveraged Random Peer Sampling (RPS) protocols to construct dynamic communication topologies—thereby speeding up convergence—we demonstrate that these systems are susceptible to a dual vector of attacks. Specifically, malicious (Byzantine) actors can inject poisoned models and manipulate the peer sampling process to disproportionately increase their impact on the network.

To counter this combined threat, we introduce GRANITE, a robust learning framework designed for sparse, dynamic graphs populated by Byzantine nodes. GRANITE tracks the identifiers of nodes it encounters over time and dynamically calibrates local aggregation thresholds according to the estimated density of Byzantine agents in each node's immediate vicinity. Our analysis reveals that, within the GRANITE framework, the concentration of Byzantine nodes in local neighborhoods decays exponentially. Additionally, we establish the specific robustness conditions for the graphs generated by this approach.

Empirical evaluations show that GRANITE achieves convergence within 5% of the accuracy attained in purely non-Byzantine settings, even when 30% of the nodes are malicious. Furthermore, the framework delivers accelerated convergence rates and reduces communication costs by a factor of up to nine compared to existing methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC