Medical AI models can leak sensitive patient information through privacy attacks, even when average risk metrics look acceptable. This paper shows that while aggregate membership inference attack success often appears close to random guessing, individual patient-level risks can be extremely high— with some patients facing near-perfect identifiability.

The risks are also disparate: they disproportionately affect underrepresented groups (by race, disease prevalence, or insurance status). Larger models tend to increase these individual and group-level privacy vulnerabilities.

AI-powered wearables — pins, pendants, glasses — are turning into always-on audio-visual recorders that capture private conversations and moments without consent.

This has sparked an escalating arms race. Early countermeasures like ultrasonic microphone jammers are already defeated by advanced AI speech-recovery models that reconstruct clean audio from noisy environments. Newer tactics shift from jamming to obfuscation: flooding systems with junk data, layered babble tapes, or real-time “anti-speech” ultrasonic signals designed to confuse reconstruction algorithms.