Understanding privacy risks in typical deep learning models for medical image analysis
Bibliographic record
Abstract
Deep learning in medical imaging typically requires sensitive and confidential patient data for model training. Recent research in computer vision has shown that it is possible to recover training data from trained models using model inversion techniques. In this paper, we investigate the degree to which encoder-decoder like architectures (U-Nets, etc) commonly used in medical imaging are vulnerable to simple model inversion attacks. Utilising a database consisting of 20 MRI datasets from acute ischemic stroke patients, we trained an autoencoder model for image reconstruction and a U-Net model for lesion segmentation. In the second step, model inversion decoders were developed and trained to reconstruct the original MRIs from the low dimensional representation of the trained autoencoder and the U-Net model. The inversion decoders were trained using 24 independent MRI datasets of acute stroke patients not used for training of the original models. Skull-stripped as well as the full original datasets including the skull and other non-brain tissues were used for model training and evaluation. The results show that the trained inversion decoder can be used to reconstruct training datasets after skull stripping given the latent space of the autoencoder trained for image reconstruction (mean correlation coefficient= 0.49), while it was not possible to fully reconstruct the original image used for training of a segmentation task UNet (mean correlation coefficient=0.18). These results are further supported by the structural similarity index measure (SSIM) scores, which show a mean SSIM score of 0.51± 0.14 for the autoencoder trained for image reconstruction, while the average SSIM score for the U-Net trained for the lesion segmentation task was 0.28±0.12. The same experiments were then conducted on the same images but without skull stripping. In this case, the U-Net trained for segmentation shows significantly worse results, while the autoencoder trained for image reconstruction is not affected. Our results suggest that an autoencoder model trained for image compression can be inverted with high accuracy while this is much harder to achieve for a U-Net trained for lesion segmentation.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".