Accountable AI with ZKPs - Certifying Fairness and Explanations under Model Confidentiality

Explore how zero-knowledge proofs can enable trustworthy AI deployment while maintaining model confidentiality in this 29-minute conference talk. Learn about the fundamental challenge of proving key properties of machine learning models without revealing the models themselves, particularly in high-stakes societal applications where legal and IP constraints require confidentiality. Discover FairProof, a system designed to publicly certify individual fairness in neural networks while preserving model secrecy, and examine ExpProof, which operationalizes explanations even in adversarial environments. Understand how these zero-knowledge proof-based approaches address the tension between transparency requirements for responsible AI and the practical need for model confidentiality, advancing the development of verifiable and accountable artificial intelligence systems.