AGILE: A Multi-task Contrastive Learning Framework with Adversarial Gradient Iterative Learning for Bio-signal Anonymization

Tamonash Bhattacharyya, Farshad Firouzi, Amir Rahmani, Sanaz Rahimi Moosavi, Krishnendu Chakrabarty

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Bio-signal Anonymization, ECG, Privacy Attack, Transformers

Abstract: Bio-signals have become a de-facto method for identity verification and subject re-identification, while providing crucial clinically relevant information for telediagnosis, thereby raising significant privacy concerns. Existing anonymization methods often operate on limited modalities leaving parts of the spatio-temporal signal space exposed to re-identification attacks while degrading signal fidelity through indiscriminate noise. Thus, we proposes a multi-task contrastive learning framework that jointly suppresses biometric features while preserving clini- cally relevant characteristics. The framework iteratively perturbs the signal using a novel adversarial fast gradient sign method (A- FGSM) for targeted noise injection that maximizes identity loss while minimizing diagnostic loss. Evaluated on PTB, CODE-15%, and MIMIC-IV-ECG datasets, our method reduces biometric identification to 15.74% while maintaining a clinical classification accuracy of 94.7%, establishing a new benchmark for bio-signal anonymization.

Track: 10. Security, privacy, and trust in digital health technologies

NominateReviewer: [e-mail]

Submission Number: 146