Learning from Encoded Patterns: A Recurrence Plot Approach for Privacy-Preserving Record Linkage

Abstract

Privacy-Preserving Record Linkage (PPRL) enables the integration of records referring to the same entities across multiple datasets without disclosing sensitive information. Traditional PPRL techniques often rely on predefined similarity metrics that may not capture complex data relationships (e.g., non-linear relationships), limiting linkage quality. This paper presents a Cross Recurrence Plot (CRP)-based representation of Bloom Filter pairs to train machine learning and deep learning classifiers capable of distinguishing matching and non-matching records directly from encoded bit patterns. The proposed workflow simplifies the PPRL process by replacing similarity thresholding with an ML-based decision model while maintaining privacy guarantees. Experimental results using real-world datasets demonstrate that the proposed approach outperforms traditional threshold-based methods in terms of linkage quality, confirming its potential to support trusted and privacy-aware data sharing.