Robust Multi-watermark Algorithm for Medical Images Based on SqueezeNet Transfer Learning
Pengju Zhang, Jingbing Li, Uzair Aslam Bhatti · 2023
The rise of intelligent healthcare has led to an increasing amount of medical data production and transmission. With the growing concern over medical data privacy, protecting sensitive information has become increasingly important. Thus, the medical image watermark technology emerged to address this issue. However, there have been few reports on the application of deep learning in medical image watermark research, and the resistance to geometric attacks is a challenging problem for robust watermark technology. To address these issues, this chapter proposes a robust multi-watermark algorithm for medical images based on SqueezeNet transfer learning. Specifically, we added a five-classification fully connected layer to the end of the SqueezeNet network structure and trained the network on a self-built enhanced medical data set. Through this approach, we can enhance the feature extraction capability of the network during transfer learning for the target-domain medical images and achieve a validation accuracy of 99.20%. This algorithm utilizes the convolutional layer of the trained network as a feature extractor and combines DCT transform to generate a zero watermark. In addition, multi-watermark technology is used to increase watermark information payload. Experimental results show that this algorithm has a certain robustness against conventional attacks. Furthermore, it has good resistance to geometric attacks and a certain degree of generalization ability.