100 Iteration (left) & 500 Iteration (right)
The paper "CT-GAN: Malicious Tampering of 3D Medical Imagery using Deep Learning" presented that the possibility of cyber attach on medicla records which leads to a serious damage on medical diagnosis process. This paper simulates the hypothetical attack by creating deep-fake X-ray images (Source: Chest X-ray Images). The ResNet-18 model, which achieved 0.98 accuracy on the origianl dataset (performance on teritary classification), performned much worse by 0.63, 0.57 and 0.55 (100, 200, and 500 iterations) on the fake images created by GAN (Generative adversarial networks). The result reiterate the danger of deep learning-based methods.
- ResNet-18 model has trained with 20 epochs and the performance in binary task (Pneumonia detection) is as follows.
| Metric | ResNet-18 |
|---|---|
| Acc(val) | 0.98 |
| Acc(test) | 0.98 |
| AUC | 0.98 |
| F1 | 1.00 |
- The model applied to perform the same task for images generated by the GAN (Generative adversarial networks).
| Metric | Iter100 | Iter200 | Iter500 |
|---|---|---|---|
| Accuracy | 0.63 | 0.57 | 0.55 |
- Iter100 = 100 iteration of epochs
- Iter200 = 200 iteration of epochs
- Iter500 = 500 iteration of epochs
- ResNet-18 model performance dropped significantly with the fake data generated by the GAN (Generative adversarial networks).
- Yisroel Mirsky, Tom Mahler, Ilan Shelef, Yuval Elovici (2019) "CT-GAN: Malicious Tampering of 3D Medical Imagery using Deep Learning", 28th USENIX Security Symposium (USENIX Security 2019). paper link
- Kermany, Daniel; Zhang, Kang; Goldbaum, Michael (2018), “Labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images for Classification”, Mendeley Data, V2, doi: 10.17632/rscbjbr9sj.2. data link