Poster Abstract: Protecting User Data Privacy with Adversarial Perturbations.

Area of Science:

• Biomedical Informatics
• Data Science
• Cybersecurity

Background:

• On-body sensors generate extensive health time-series data, enabling advanced data-driven models for well-being inference.
• Sharing this sensitive data facilitates cross-institutional collaboration but raises significant privacy concerns, particularly regarding membership inference attacks.

Purpose of the Study:

• To develop a method for protecting clinical time-series data against membership inference attacks.
• To ensure maximal retention of data utility while enhancing privacy.

Main Methods:

• Implementing adversarial perturbations, which involve adding specially trained, imperceptible noise to raw time-series data.
• Training noise to induce inference mistakes in deep learning models, specifically targeting user identity prediction.

Main Results:

• The proposed solution demonstrates superior protection against membership inference attacks compared to existing baseline methods.
• The method successfully passed all designed data quality checks, indicating preserved data utility.

Conclusions:

• Adversarial perturbations offer an effective strategy for safeguarding privacy in clinical time-series data.
• This approach balances robust privacy protection with the critical need for data utility in health research.