Active Fourier Verifier: PAC Estimation of Model Properties with Influence Functions and Fourier Representations

With the increasing deployment of Machine Learning (ML) models in real-world applications, verifying and auditing properties of ML models have become a central concern. In this work, we focus on three properties: group fairness, individual fairness, and robustness. Based on Boolean function analysis, we propose an influence function-based framework to compute these three properties for ML models with binary input. We further leverage Fourier expansion-based techniques to compute them accurately and efficiently. We provide high probability error bounds on the proposed estimates. We also extend the proposed approach to categorical and continuous input. Finally, we numerically evaluate the accuracy of the proposed algorithm to estimate group fairness, individual fairness, and robustness of different ML models.