Enhanced Adversarial Drift Detection for MLOps Feature Monitoring

Machine learning models in production suffer from distribution drift, where evolving input features degrade performance. Existing unsupervised detectors signal that drift occurred, but not which features drifted or how severe the shift is. This paper proposes Enhanced Adversarial Drift Detection (EADD), extending the D3 discriminative detector with three contributions: (1) an Adaptive Sequential Permutation Test (ASPT) for rigorous drift confirmation, cutting permutation budget by 90-95%; (2) SHAP-based drift localisation and a Drift Severity Index (DSI) for feature-level diagnosis and severity quantification; and (3) benchmarking against seven unsupervised detectors on synthetic and real-world streams. EADD achieves full drift type coverage (4/4) with zero false alarms on synthetic data, detects drift on three of four annotated INSECTS datasets, and outperforms D3 on detection delay in all three detected cases (up to 8× faster). SHAP attribution reaches perfect accuracy (P@k = R@k = 1.0) on controlled synthetic scenarios. On autocorrelated streams, EADD produces zero false alarms versus D3's 54.6 (_p =0_.0101, Mann-Whitney U).