Where you place a few bad actors in a network changes how badly they can break the whole system.

In decentralized federated learningdecentralized federated learningA way for devices to train a shared model without central coordination, using peer-to-peer communication., many devices train an AI model together by sharing updates directly with their neighbors across a network. Bad actors, called Byzantine participants, can try to corrupt this shared model. Most security studies look only at how these attackers behave, not which specific devices they take over. Because devices pass messages from neighbor to neighbor, where attackers sit completely changes how their bad updates spread. Researchers treated choosing which nodes to corrupt as a key attacker strategy under a fixed budget. To find the worst spots without running slow simulations, they created a metric called Byzantine Placement Influence. It tracks how bad updates spread over multiple network hops over time. Their tests showed this method consistently finds the most damaging attack positions, proving that attacker location is vital for network defense.

Why it matters

This affects the security of AI systems where many devices train models together without a central server. If attackers can choose which devices to compromise, they can cause more damage to the shared model, impacting users who rely on it.

Questions answered