Xiuli Chai

In the context of the Internet of Medical Things (IoMT), medical images are highly vulnerable to unauthorized analysis and recognition by deep neural networks (DNNs) during transmission and sharing, posing serious threats to patient privacy and the rights of legitimate users. Reversible Adversarial Examples (RAEs) have been proposed to prevent unauthorized DNNs from interpreting image content without affecting normal usage by authorized parties. However, existing RAE methods primarily focus on privacy protection while largely neglecting copyright issues. To address this issue, we propose a robust RAE generation algorithm based on digital watermarking, which achieves copyright protection and reversible adversarial protection. Specifically, LayerCAM is used to determine the watermark embedding region (WR), and a multi-objective optimization method with dynamic weight adjustment named as DW-NSGA-PSO is presented to optimize the embedding position and transparency. Then, the alpha blending strategy is adopted to fuse the watermark into the original sample, generating a watermarked adversarial example. Next, the watermarked image with location information serves as the base sample, and optimized adversarial perturbations are embedded into the WR to resist removal attack of watermark removal networks. Finally, reversible data hiding (RDH) is used to embed information about the WR into the non-embedding regions of the image, enabling reversible adversarial protection. Besides, our algorithm further enhances the transfer attack capability of RAEs and robustness against various image attacks by randomly preprocessing the images. Experimental results demonstrate that the proposed algorithm achieves significantly attack performance on both color and grayscale image datasets, reaching a maximum ASR of 91%, and it may effectively protect both privacy and copyright across multiple datasets while maintaining good visual quality.

The popularity of wearable medical data collection and surveillance devices provides real-time guarantees for the whole process of a patient's medical treatment, especially medical image data plays a key role. However, existing medical images face data leakage, pollution and vulnerability to attacks during transmission over wireless body area networks(WBANs). To address these issues, a privacy protection algorithm based on Hopfield cross neural network (HCNN) for medical data is proposed. Specifically, the HCNN model is first constructed and its dynamic behavior is analyzed, which is suitable for application to image encryption. Then, a confusion method of NZ fractal curve sorting matrix (NZ-FCSM) is designed to achieve good encryption effect. Subsequently, the secret image sharing (SIS) technique based on sharing matrix is introduced to enhance the algorithm robustness. Finally, an alignment embedding of double diamond prediction (AEDDP) method is proposed to implement lossless hiding of private information. The present issues in medical image protection include ensuring the security and effectiveness of encryption algorithms while maintaining the robustness and concealment of ciphertext data, and balancing the need for preservation with the limited resources of complex work environment. Experimental results show that the proposed algorithm achieves PSNR of 53 dB for the cipher image, more than 36 dB for the reconstructed image, and the information entropy of the secret image is over 7.99, and displays good robustness. These findings highlight the validity of the algorithm in medical image data privacy preserving applications that ensure confidentiality and extend to practical applications of concealed transmission of confidential information and secure multi-party transactions.

Adversarial example detection plays an important role in defending deep neural networks (DNNs) from assaults by distinguishing whether the inputs are benign or manipulated. However, current detection algorithms leave much to be improved in terms of accuracy and stability for high-resolution and slightly perturbation adversarial examples. To address this issue, we propose a multi-information interaction enhancement model based on pixel-feature space for adversarial example detection. Specifically, we design the enhancement module based on statistical information and FFT to extract and enhance feature activation, which contains the most direct evidence for discriminating adversarial perturbations. Besides, we interactively fuse the pixel stream and feature stream data over the channel to obtain dual stream data. Eventually, we devise and add channel variance-pooling in the spatial attention unit to further improve the model’s search for adversarial perturbations. To fully validate the performance of the model, we conduct extensive experiments and analyses on five types of mainstream adversarial examples from the CIFAR-10 and ImageNet datasets. Average F1 scores of 0.95 and 0.91 are achieved on the two datasets, respectively, which proves that the proposed method can accurately and stably detect multiple adversarial examples and effectively alleviate the negative impact of adversarial attacks on DNNs.