SPADR: A Context-Aware Pipeline for Privacy Risk Detection in Text Data

Large language models (LLMs) are powerful, but they can unintentionally memorize and leak sensitive information found in their training or input data. To address this issue, we propose SPADR, a semantic privacy anomaly detection and remediation pipeline designed to detect and remove privacy risks from text. SPADR addresses limitations in existing redaction methods by identifying deeper forms of sensitive content, including implied relationships, contextual clues, and non-standard identifiers that traditional NER systems often overlook. SPADR combines semantic anomaly scoring using a denoising autoencoder with named entity recognition and graph-based analysis to detect both direct and hidden privacy risks. It is flexible enough to work on both training data (to prevent memorization) and user input (to prevent leakage at inference time). We evaluate SPADR on the Enron Email Dataset, where it significantly reduces document-level privacy leakage while maintaining strong semantic utility. The enhanced version, SPADR (S2), reduces the PII leak rate from 100% to 16.06% and achieves a BERTScore F1 of 88.03%. Compared to standard NER-based redaction systems, SPADR offers more accurate and context-aware privacy protection. This work highlights the importance of semantic and structural understanding in building safer, privacy-respecting AI systems.