KNT: Lightweight Privacy Layer Cuts Patient Leakage 36%
New technique reduces re-identification risk by 36% with only 0.15ms overhead.
A new paper on arXiv from researchers Haebom Lee and Gyeongjung Kim introduces Keyed Nonlinear Transform (KNT), a lightweight privacy-enhancing technique for medical image analysis. The method addresses a critical vulnerability in split inference, where hospitals share intermediate feature representations with cloud servers. These features can inadvertently leak patient identity information. KNT applies a key-conditioned nonlinear obfuscation to intermediate representations, reducing re-identification AUC from 0.635 to 0.586—a 36% reduction in above-chance identity signal. The transform introduces only 0.15 ms CPU overhead and requires no backbone retraining, while preserving classification performance within 1.0 percentage point. Additionally, KNT prevents closed-form inversion and forces any attempted recovery into iterative gradient-based optimization, even under full key compromise.
Beyond classification, KNT generalizes to dense prediction tasks like skin-lesion segmentation, incurring only a 4.4 pp reduction in Dice score without retraining. This positions KNT as a practical and efficient privacy layer for split inference deployments in resource-constrained hospitals. The technique allows healthcare institutions to leverage cloud-based AI for medical image analysis without exposing sensitive patient data, potentially accelerating adoption of AI diagnostics while maintaining regulatory compliance.
- Re-identification AUC reduced from 0.635 to 0.586, a 36% reduction in above-chance identity signal
- Only 0.15 ms CPU overhead and no backbone retraining required, preserving 1.0 pp classification accuracy
- Generalizes to segmentation tasks with just 4.4 pp Dice reduction, applicable without retraining
Why It Matters
Hospitals can now securely share medical images with cloud AI without leaking patient identities, enabling privacy-compliant diagnostics.