The Unseen Hand: Manipulating Model Fairness and SHAP with Targeted Identity Re-Association Attacks

The article introduces Targeted Identity Re-Association (TIRA) attacks, a novel method for manipulating algorithmic fairness and explainability in machine learning models. It details two algorithms: Probabilistic Micro-Shuffling (PMiS) and Probabilistic Rank-Shift Micro-Perturbation (PRSMP), which can subtly alter model outputs without requiring internal access, effectively skewing fairness metrics and confounding SHAP-based explanations. This research highlights significant vulnerabilities in model auditing mechanisms, emphasizing the need for enhanced robustness in fairness and explainability tools for AI practitioners.