Deploying Trustworthy AI - From traceability, robustness and interpretability

Artificial intelligence is increasingly being deployed in high-stakes domains, including energy, transportation, healthcare, and industrial operations. While foundation models and machine learning systems demonstrate remarkable capabilities, their widespread adoption remains constrained by concerns regarding reliability, transparency, accountability, and human oversight. Ensuring that AI systems behave as intended under real-world conditions has become a central challenge for both research and practice.

In this talk, I present a comprehensive perspective on trustworthy AI through three complementary dimensions: traceability, robustness, and interpretability. First, I introduce the concept of a Decision Bill of Materials (DBOM), a tamper-resistant framework that extends software supply-chain transparency to AI training and inference pipelines through trusted computing technologies. Second, I discuss recent advances in evaluating and improving robustness against adversarial manipulations, distribution shifts, and uncertainty, including work on adversarial attacks and conformal prediction-based reliability guarantees. Third, I examine the role of explainability in enabling effective human oversight, highlighting both the limitations of existing post-hoc explanation methods and recent developments in concept-based interpretable learning.

Drawing on research at the intersection of machine learning, security, and dependable systems, this talk argues that trustworthy AI cannot be achieved through any single technique. Rather, deployable AI requires the integration of traceable decision processes, robust behavior under uncertainty, and meaningful explanations that support human decision-making. Together, these capabilities provide a foundation for building dependable AI systems suitable for safety-critical industrial environments and future foundation-model-based applications.