Exposing the Illusion of Fairness: Auditing Vulnerabilities to Distributional Manipulation Attacks
A new research paper explores how malicious actors can manipulate AI fairness audits to create an illusion of compliance. The study, published on arXiv, details strategies for constructing manipulated datasets that appear representative while violating fairness constraints, particularly concerning the EU AI Act's high-risk classifications. Researchers propose statistical tests based on distributional distance to detect these manipulations and provide guidelines for strengthening verification processes. AI
IMPACT Highlights potential vulnerabilities in AI fairness auditing, urging stronger verification methods to prevent deceptive compliance.