Beyond Trust - Proving Fairness and Privacy in Machine Learning

Explore advanced techniques for verifying fairness and privacy claims in machine learning through this Google TechTalk by Ali Shahin Shamsabadi. Examine the limitations of current post hoc auditing methods that allow institutions to engage in "FairWashing" or apply "privacy perfume" without genuine implementation of fair and private training practices. Learn about Zero Knowledge Proof (ZKP) frameworks that enable white-box auditing while maintaining confidentiality, including Confidential-PROFITT for proving fair training of decision trees and Confidential-DPproof for demonstrating differentially private training. Discover how these innovative approaches allow institutions to provide verifiable proof of their privacy-preserving and fairness-aware machine learning practices without compromising the confidentiality of their data and models. Gain insights into the research behind identifying and mitigating AI system failure modes, and understand the development of confidential auditing frameworks that move beyond trust-based systems to mathematically verifiable proof systems.