Ensuring Resilience and Efficiency in Distributed AI Systems

The shift toward physical AI, IoT and on-device agents, combined with privacy, latency and bandwidth constraints, is driving the need for distributed intelligence at the edge. Distributed AI unlocks siloed data and enables collaborative learning across devices while keeping data local. However, it faces challenges from limited and heterogeneous edge resources (data, computation and connectivity), the need to learn continually from streaming data, and new security and privacy threats. This talk gives an overview of our research on making distributed and federated AI both efficient and trustworthy, spanning resource-efficient and continual federated learning, hyperdimensional computing, and adversarial robustness and privacy. It also highlights our open- source tools: OpenFL, MART, LLMart and ATLAS. The talk closes with future directions for securing emerging multi-agent and distributed AI systems.