Assessing the Security of Personal Data in Large Scale chest X-Ray Image Screening

This paper is related to the problem of security of personal data in the context of massive screening of the population. We attempt to assess the probability of identification of particular individuals by their chest images stored together with other private data in large-scale image databases. It was supposed that the attacker have X-Ray image of target subject but taken several years earlier at different scanning conditions and uses it as a key. A total of 90,000 images of 45,000 subjects were sampled from a database containing 1,909,000 records. The study groups were fully balanced by both age and gender. Image features were derived using 3 different CNNs including EfficientNet-B0, EfficientNet-B0-V2, and BiT-S R50x1. Results of searching correct image of a pair for all 45,000 people were presented in form of the fraction of correct answers in the Top-N most similar while N runs from 1 (correct answer on the first position) up to 80. It was found that (a) EfficientNet-B0 produces the best image features among the three CNNs being examined. (b) The fraction of correct identifications of subjects that is the right answers appeared on the first position was about 14% whereas the percentage of correct results in Top-80 achieved 33%. (c) The chances to be identified are significantly higher in female subjects compared to males and higher in young subjects compared to the aged ones