Workshops

As AI becomes part of everyday life, security failures in machine learning directly affect people’s safety and well-being, yet few communities specifically focus on how technical ML vulnerabilities lead to real-world harm or which protections are truly needed. To answer these questions, the Human-Centered Trustworthy Machine Learning workshop brings together ML security researchers, human-subjects experts, and others interested in improving the security and privacy of ML systems in real-world settings. We encourage work that grounds technical defenses in practical needs, or improves the methods for evaluating the real-world impact of ML systems. We aim to strengthen interdisciplinary collaboration across technical and human-centric work, and welcome anyone eager to learn and contribute to human-centric trustworthy ML systems.

The MIRROR workshop aims to critically reflect on research practices in AI Security and explore how our community can conduct more reliable and reproducible research. Building better research practices requires looking in three directions at once. Looking backward means examining where current research goes wrong: brittle threat models and insufficiently stress-tested defenses can create a false sense of security, making systematic audits of pitfalls and lessons learned across communities and subfields essential. Looking forward means embedding sound evaluation protocols, reproducible artifacts, data and label quality, and responsible disclosure into research practice. Looking inward matters just as much: extracting signal from a growing volume of papers and communicating ideas clearly is increasingly important, especially as LLMs reshape ideation, writing, and peer review.

Collaborative machine learning only succeeds if participation is worthwhile. WICE is the first workshop at a security venue dedicated to contribution evaluation and incentives in collaborative learning: Shapley-based data valuation, manipulation-resistant and privacy-compatible contribution scores, free-rider and misbehavior detection, incentive mechanisms, and credit assignment in multi-agent systems. The half-day program features an invited keynote, a tutorial, contributed short papers (published in the proceedings), and a panel with academia and industry. Paper submission deadline: December 18, 2026.

TrustTFM is the first workshop dedicated to the trustworthiness of Tabular Foundation Models (TFMs), an emerging class of models for structured data. Tabular data drives decision-making in finance, cybersecurity, healthcare, energy, and other regulated or safety-critical domains, and regulations such as the EU AI Act and the NIST AI Risk Management Framework now make demonstrable trustworthiness a legal prerequisite rather than a technical aspiration. As TFMs move toward deployment, reliable ways to assess their robustness, security, privacy, explainability, fairness, and compliance risks are essential. The workshop brings together researchers and practitioners to develop realistic threat models, benchmarks, and evaluation methodologies that support the safe adoption of TFMs in real-world systems.