The Contrastive Credibility Propagation Algorithm in Action: Improving ML-powered Data Loss Prevention

This report presents the Contrastive Credibility Propagation (CCP) algorithm, a semi-supervised learning method that iteratively refines pseudo-labels using credibility vectors and optional subsampling to improve robustness to common real-world data issues (e.g., few labels, open-set noise, label errors, and class imbalance). It outlines the architecture and a softly supervised contrastive loss, reports consistent performance against supervised and SSL baselines across varied conditions, and demonstrates a Data Loss Prevention (DLP) use case that leverages differentially private, unlabeled production data to align models with deployment distributions, yielding substantial real-world detection improvements.