Consensus-based algorithms find various applications in distributed multi-agent systems. In this talk we focus on two such applications and discuss two novel algorithms that have been developed as part of the PhD research work.
Firstly, we address dynamic average consensus (DAC) in discrete-time. The challenge here is the presence of Byzantine adversarial attacks. To ensure resilience a subset of the agents, termed trusted agents, are assumed to have higher security guarantees. We first formally define resilient DAC in the presence of Byzantine adversaries. Then, we present our novel ResDAC algorithm, that ensures the trusted and ordinary agents achieve resilient DAC in the presence of adversarial agents.
Then we move to federated learning in a deterministic setting. While most existing literature considers a single central server for all the clients, we propose a distributed federated learning system with multiple servers and different sets of clients associated with each server. We present our novel DFL algorithm which ensures that all the servers eventually agree on a common parameter model and this model lies within a small tolerance from the ideal model.