Reproduce “Model-Agnostic Utility-Preserving Biometric Information Anonymization” on OpenResearch

About this paper

Chun-Fu Chen, Bill Moriarty, Shaohan Hu, Sean Moran, Marco Pistoia, Vincenzo Piuri, Pierangela Samarati

A model-agnostic framework, developed by researchers at JPMorgan Chase Bank, N.A. and Università degli Studi di Milano, transforms biometric feature vectors via selective weighted averaging within dynamically assembled, attribute-pure sets to anonymize data while preserving utility for machine learning. The approach achieved high recognition accuracy for attributes of interest (e.g., 78.8% for sentiment, 93.2% for spoken digits) and strong suppression of sensitive attributes (mixture levels near 1.0, low KL divergence) across facial, voice, and motion biometric datasets.

Key insights

• Selective feature weighting within dynamically assembled, attribute-pure groups allows for fine-grained control over the utility-privacy trade-off in high-dimensional biometric data.
• The purity parameter for assembling random sets directly influences the preservation of the attribute of interest's utility, while the set size primarily determines the degree of sensitive attribute mixture.
• Actively suppressing features relevant to sensitive attributes during the feature selection process dramatically improves anonymization quality without significant loss of utility, especially at higher feature retention ratios.

Summary by alphaXiv.