Upcoming Seminars

Abstract: Multi-agent reinforcement learning (MARL) has emerged as a promising approach for adaptive and scalable cyber defense, enabling distributed agents to learn coordinated defense policies in complex and uncertain network environments. Its ability to adaptively optimize decisions across multiple agents makes MARL appealing for defending against evolving cyber threats. In a broader area of artificial intelligence, existing adversarial examples have shown that even sophisticated models can be misled, causing classifiers or learned policies to make incorrect decisions. This raises important questions for cyber defense: could similar adversarial strategies undermine MARL-based defenders, and how resilient are they when facing intelligent and coordinated deception? 

In our work, we investigate the design of AI-powered adversaries that challenge MARL defense policies in distributed network environments. By modeling defenders’ interactions as a decentralized decision-making process under uncertainty, we develop new classes of adversarial strategies that intelligently manipulate feedback, disrupt information flow, and interfere with coordination—operating under both resource and stealth constraints to systematically degrade groups of AI decision makers. Our findings reveal critical gaps in MARL defense mechanisms and motivate next-generation security frameworks that explicitly account for deception-aware adversaries, advancing the robustness of AI-driven cyber defense in dynamic and adversarial environments.

Abstract: This talk explores how principles of cyber-physical system (CPS) security and resilience can be effectively translated into the design of secure, autonomous, and self-healing microservices within the FinTech sector. Drawing from real-world industry experience, I will demonstrate how agentic AI, performance-aware microservices, and zero-trust principles can be used to safeguard high-volume financial transactions from disruptions, anomalies, and threats. The session will cover practical architectures for self-tuning systems, discuss failure domains in distributed transaction systems, and present secure-by-design strategies for cloud-native applications. By aligning industry-grade systems with CPS-like resilience models, this talk aims to foster discussion between theoretical control systems security and its real-world application in modern digital finance infrastructure.

Abstract: Modern societies increasingly rely on AI-driven cyber-physical-human systems (CPHSs), such as intelligent transportation, industrial automation, and other critical infrastructure. While these systems promise efficiency and intelligence, they also introduce new vulnerabilities where security, privacy, and resilience are tightly coupled with human trust. A central question arises: how can we design socio-technical systems that remain trustworthy and resilient even in the presence of adversarial manipulation and the cognitive biases of human decision-makers? In this talk, we will present a research agenda that develops principled and computationally tractable frameworks for understanding trust in CPHSs. We will walk through four key perspectives: assessing trust via meta-game analysis of human–CPS interactions, building trust in AI through crowd auditing and accountability mechanisms, exploiting trust in adversaries through defensive deception, and maintaining user trust under misinformation with information design strategies. These frameworks will be illustrated through case studies in critical CPHS domains. We will conclude by outlining future directions toward resilient, cognitive-aware CPHSs.

Abstract: Bayesian data fusion offers a principled route to distributed learning under privacy and uncertainty. This talk develops a unifying framework that clarifies how local beliefs should be combined when priors are shared. We analyze the Conditionally Independent Likelihood (CIL) and Conditionally Independent Posterior (CIP) rules, identify the prior double-counting pitfall in naïve posterior multiplication, and derive corrections that preserve coherence while characterizing accuracy as a function of client count and prior informativeness, beyond Gaussian models. Building on this foundation, we introduce federated posterior sharing for multi-agent systems, in which agents exchange posteriors rather than data to construct a global belief and act. The method supports single-shot or periodic synchronization, avoids prior reuse, and improves reward and sample efficiency under uncertainty and heterogeneity. Finally, we present a Bayesian formulation of clustered federated learning that treats client–cluster assignment as latent data association, yielding practical approximations that handle non-IID feature and label skew and outperform standard clustered FL. Together, these results provide a coherent recipe—fuse beliefs, correct for shared priors, and quantify uncertainty—for privacy-preserving learning and decision making at scale.