Medical imaging is increasingly shared as learned embeddings rather than raw scans, on the assumption that removing the original images protects patient privacy. A new study tests that assumption for brain MRI by probing five frozen pretrained encoders on two longitudinal T1-weighted datasets, OASIS-2 and MIRIAD.

The authors simulated an attacker who has access to released embeddings, the corresponding public encoder, and auxiliary MRI scans. Across all representations, they found substantial subject-specific information: the highest-risk encoder allowed rank-1 patient linkage in up to 99% of cases on MIRIAD and over 90% on OASIS-2. The embeddings also leaked a demographic attribute, with recorded-sex prediction reaching balanced accuracy of 90.9% on OASIS-2 and 95.5% on MIRIAD.

These results hold even though none of the encoders was trained with an identification objective. The authors conclude that replacing raw brain MRI with learned embeddings does not, by itself, eliminate patient-linkage or attribute-inference risk, and is insufficient on its own to establish anonymity.