Trust and Interaction-Driven Resilient Distributed Secure Learning Algorithm and Its Application

This paper investigates the problem of distributed non-convex secure learning for multi-agent systems in non-secure and non-trusted environments, where only local datasets are available. To enhance the system's security and privacy during information exchange, we propose a trust and interaction-driven robust adaptive learning algorithm. Theoretical analysis demonstrates that the proposed algorithm effectively protects agent privacy while mitigating the impact of various types and arbitrary numbers of non-cooperative agents, without incurring additional communication or computational overhead. Finally, experiments on the image recognition task validate the algorithm's performance, showing that it outperforms existing methods in terms of security and privacy.