An Analysis of the Vulnerability of Two Common Deep Learning-Based Medical Image Segmentation Techniques to Model Inversion Attacks
Bibliographic record
Abstract
Recent research in computer vision has shown that original images used for training of deep learning models can be reconstructed using so-called inversion attacks. However, the feasibility of this attack type has not been investigated for complex 3D medical images. Thus, the aim of this study was to examine the vulnerability of deep learning techniques used in medical imaging to model inversion attacks and investigate multiple quantitative metrics to evaluate the quality of the reconstructed images. For the development and evaluation of model inversion attacks, the public LPBA40 database consisting of 40 brain MRI scans with corresponding segmentations of the gyri and deep grey matter brain structures were used to train two popular deep convolutional neural networks, namely a U-Net and SegNet, and corresponding inversion decoders. Matthews correlation coefficient, the structural similarity index measure (SSIM), and the magnitude of the deformation field resulting from non-linear registration of the original and reconstructed images were used to evaluate the reconstruction accuracy. A comparison of the similarity metrics revealed that the SSIM is best suited to evaluate the reconstruction accuray, followed closely by the magnitude of the deformation field. The quantitative evaluation of the reconstructed images revealed SSIM scores of 0.73±0.12 and 0.61±0.12 for the U-Net and the SegNet, respectively. The qualitative evaluation showed that training images can be reconstructed with some degradation due to blurring but can be correctly matched to the original images in the majority of the cases. In conclusion, the results of this study indicate that it is possible to reconstruct patient data used for training of convolutional neural networks and that the SSIM is a good metric to assess the reconstruction accuracy.
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 | 0.001 | 0.000 |
| 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) | 0.000 | 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".