A Review of Adversarial Behaviour in Distributed Multi-Agent Optimisation

摘要

This paper addresses the challenges of distributed multi-agent optimisation, in environments with potential adversarial agents. In distributed multi-agent optimisation, each agent has a local cost function and the collective goal is to optimise the global cost function, which takes as input, the output of each agent's local cost function. Such optimisation algorithms are used in different fields such as distributed machine learning, distributed robotics, and recently, distributed energy planning. A prominent assumption in most of these algorithms is that all the agents are cooperative, non-faulty, and non-adversarial. Yet, in scenarios with multi-agent systems, such assumptions are not always valid. Recently, there has been some research in the area of resilience and fault-tolerant distributed multi-agent optimisation. However, these studies either assume a fully Byzantine environment, or have their own adversary model with different sets of assumptions about adversary capabilities. This makes the analysis and comparison between the results challenging. This paper presents a review of such algorithms which helps in defining and investigating restricted adversarial behaviour in distributed multi-agent optimisation. In comparison to the Byzantine environments which carry no assumptions about the agent's capability or purpose, in this paper adversaries have restrictions in regard to their knowledge, capabilities, purpose, and impact. This approach makes it easier to provide guarantees about the system performance depending on the specific adversary model and avoid the overly conservative, 'one size fits all' algorithms needed in fully Byzantine environments. Moreover, using the gained insights, previously unexplored aspects of adversarial behaviour in distributed multi-agent optimisation are discovered and indicated as possible future research directions.