InfoShield: Privacy-Preserving Speech Representations for Mental Health Screening via Information-Theoretic Optimization

Researchers introduce InfoShield, a privacy-preserving machine learning technique that maintains depression detection accuracy while preventing the inference of sensitive demographic attributes from speech data. The method uses information-theoretic optimization to reduce mutual information between speech representations and demographic information, addressing a critical barrier to clinical deployment of speech-based mental health screening.

InfoShield addresses a fundamental tension in healthcare AI: the need for effective diagnostic tools versus patient privacy expectations. Mental health screening through speech analysis shows clinical promise for scalable depression detection, yet healthcare providers hesitate deploying such systems due to legitimate concerns that demographic data could be extracted from voice patterns, creating privacy risks for vulnerable populations. The research tackles this directly by introducing TimeAwareMINE, an innovation that improves mutual information estimation for sequential data through cross-modal attention mechanisms.

This work emerges from broader challenges in privacy-preserving machine learning. Traditional approaches like adversarial training prove brittle against novel attacks, while differential privacy's noise injection degrades diagnostic performance substantially. InfoShield's information-theoretic approach offers a middle ground, reducing gender inference accuracy from 92.6% to 55.5% and age inference from 55.7% to 30.3% while maintaining strong depression detection (F1=0.784 versus prior state-of-the-art 0.723).

The implications extend across healthcare AI deployment. Mental health screening represents a significant use case where privacy concerns genuinely impede adoption. Successfully decoupling diagnostic utility from demographic leakage could enable broader clinical deployment of speech-based diagnostics, benefiting telemedicine platforms and global mental health accessibility. The technical contribution—improving MINE estimators for temporal data—has applications beyond mental health in any domain requiring privacy-preserving speech analysis.

Looking forward, validation across additional datasets and diverse populations remains critical. The 6% F1 reduction versus unrestricted models represents acceptable performance loss for many clinical settings, but deployment success depends on regulatory acceptance and clinical validation studies demonstrating that privacy protections don't introduce systematic diagnostic biases across demographic groups.

• →InfoShield reduces demographic attribute inference from speech while preserving depression detection accuracy through information-theoretic optimization.
• →TimeAwareMINE improves mutual information estimation for sequential speech data using cross-modal attention, addressing temporal-static alignment challenges.
• →Gender and age inference rates drop substantially (92.6% to 55.5% and 55.7% to 30.3% respectively) with only 6% F1 performance loss.
• →The approach bridges privacy-accuracy tradeoffs that limit clinical deployment of speech-based mental health screening systems.
• →Technical innovations in privacy-preserving ML for speech data could enable broader adoption of telemedicine and remote mental health diagnostics.