Optimizing Byzantine Node Placement in Decentralized Federated Learning
A research paper introduces Byzantine Placement Influence (BPI), a set-level measure derived from gossip dynamics to quantify cumulative exposure of honest nodes to Byzantine sources. It formulates Byzantine placement as an adversarial decision under a fixed compromise budget and develops efficient algorithms for optimizing BPI, evaluated across six heterogeneous graph families with untargeted model poisoning.
The paper treats Byzantine node placement as an explicit adversarial decision in decentralized federated learning. It proposes BPI to approximate the attacker's objective of maximizing finite-time impact on honest nodes without executing learning for every candidate placement. BPI accounts for weighted multi-hop propagation and interactions among compromised nodes, unlike centrality heuristics. Algorithms for optimizing BPI are developed and evaluated across six heterogeneous graph families.
BPI is a set-level measure derived from actual gossip dynamics, capturing cumulative exposure over the training horizon. It directly accounts for weighted multi-hop propagation and interactions among compromised nodes, providing a more accurate approximation of adversarial impact than node centrality heuristics. Efficient optimization algorithms are developed for this measure.
The research highlights a gap in security evaluations of decentralized federated learning, which typically overlook which participants are compromised. By focusing on placement, it suggests that defending against Byzantine attacks requires considering network topology and propagation dynamics, not just behavior of malicious nodes.
For organizations deploying decentralized federated learning, this research offers a method to assess and mitigate risk from compromised nodes, potentially improving security posture and trust in distributed AI systems.
Future work may extend BPI to other attack types or defense mechanisms, and validate on real-world decentralized learning systems. The approach could inform robust graph design or node selection strategies to mitigate Byzantine influence.