Implicit neural obfuscation for privacy preserving medical image sharing

Despite its undeniable success, deep learning for medical imaging with large public datasets leads to an often overlooked risk of leaking sensitive patient information. A personś X-ray, even with proper anonymisation applied, can readily serve as fingerprint and would enable a highly accurate re-identification of the same individual in a large pool of scans. Common practices for reducing privacy risks involve a synthetic deterioration of image quality, e.g. by adding noise or downsampling images, before sharing them publicly. Yet, this also adversely affects the quality of downstream image recognition models trained on such datasets. We propose a novel strategy for finding a better compromise of model quality and privacy preservation by means of implicit neural obfuscation. Our method jointly overfits a neural network to a small batch of patientsX́-ray scans and applies a substantial compression - the number of network parameters representing the images is more than 6x smaller than the original images. In addition, we introduce a k-anonymity mixing that injects partial information from other patients for each reconstruction. That way identifiable information is efficiently obfuscated, while we manage to maintain the quality of relevant image parts for the intended downstream task. Experimental validation on the public RANZCR CLiP dataset demonstrates improved segmentation quality and up to 3 times reduced privacy risks compared to a more basic image obfuscation baselines. In contrast to other recent work that learn specific anonymous representations, which no longer resemble visually meaningful scans, our approach remains interpretable and is not tied to a certain downstream network. Source code and a demo dataset are available at https://github.com/mattiaspaul/neuralObfuscation.