Metrics and a Single-Pass Framework for Disentangling Feature-Level Privacy Risk in Machine Learning Models

This paper introduces a diagnostic pipeline for evaluating feature-level privacy risk in machine learning (ML) models, anchored by the Coordinate Dependency Index (CDI)—a unitless, cross-modality metric that is designed to quantify the relative degree of coordinate dependence that persists following an ML model’s compression process. Grounded in the Information Bottleneck principle, the CDI establishes an exact algebraic threshold—the Invariance Horizon at e−1≈0.37—to detect coordinate dependence across disparate data types. To facilitate direct cross-modality comparisons of the underlying leakage, we integrate Absolute Information Gain (AIG) as a complementary metric that maps these threshold exceedances into absolute bits. To operationalize this dual-metric approach, we develop the Representational Interrogator (RI), an inverse multi-task learning architecture that resolves the long-standing Audit Scalability Gap in ML privacy auditing. Experimental results across 30 trials demonstrate that the RI framework maintains strong initialization stability and achieves single-pass efficiency by collapsing traditional, sequential privacy auditing batteries into a unified training objective. Achieving speedups of up to 91× over generative adversarial batteries while maintaining strong diagnostic rank-order alignment, the framework’s robustness across deep neural networks and gradient-boosted decision trees validates the RI-CDI-AIG pipeline as a principled, architecture-independent methodology for disentangling and ranking informational vulnerabilities in ML models.