Xu An Wang

With the rapid development of the Internet of Medical Things (IoMT), wearable sensors and intelligent medical terminals continue to generate a large amount of sensitive medical data, which needs to be uploaded to the cloud platform for storage and sharing. However, IoMT devices usually have the characteristics of limited computing resources and limited energy, and it is difficult for traditional high-complexity encryption schemes to meet the requirements of security and efficiency. In addition, the existing Proxy Re-Encryption (PRE) scheme has the risk of authorization transfer, and it is difficult to achieve fine-grained and secure sharing of medical data. In response to the above problems, this paper proposes a lightweight non-pairing proxy re-encryption enhancement scheme SM2-PRE+ based on the SM2 algorithm, which is used for the secure sharing of IoMT medical sensing data. The scheme combines the SM2 elliptic curve cipher algorithm and the SM4 symmetric encryption algorithm to achieve data protection through a double-layer key structure, and it uses the re-encryption mechanism of message binding to enhance the authorization control ability. Compared with the traditional PRE scheme, the proposed scheme avoids bilinear pairing operations, reduces the computing overhead of resource-limited equipment, and supports collusion-resistant and authorization non-transferability. Security analysis shows that the scheme meets the security requirements of IND-CCA under the Random Oracle Model (ROM). Performance analysis results show that SM2-PRE+ has low computing overhead and storage burden, which is suitable for wearable medical devices, intelligent sensing terminals, and cloud-assisted IoMT data sharing scenarios.

As an increasing number of neural network models are deployed in the cloud, privacy concerns have surged, primarily due to the exposure of plaintext user data. To address the core challenge in ciphertext inference where polynomial activation functions struggle to balance model expressive capability and encrypted computation efficiency, this paper proposes a cipher-text CNN inference scheme with mutual information-optimized adaptive polynomial activation functions. Based on the CKKS homomorphic encryption algorithm, the scheme quantifies the correlation between features of each neural network layer and labels via mutual information, dynamically matching the degree and coefficients of polynomial activation functions. First, in the plaintext training phase, mutual information statistics are utilized to classify the functional importance levels of each network layer. Second, polynomial coefficients are optimized through end-to-end training, and combined with threshold filtering and numerical scaling in the ciphertext domain, ensuring the coefficients comply with the constraints of encrypted computation. Finally, a lightweight CNN architecture is integrated with the TenSEAL encryption framework to achieve end-to-end ciphertext inference. Experiments on the MNIST dataset and Fashion-MNIST dataset demonstrate that the proposed scheme achieves a plaintext model accuracy of 98.34%, 95.62% respectively and a ciphertext inference accuracy of 98.13%, 95.31% respectively. It significantly outperforms fixed-degree polynomial activation schemes, realizing the synergistic optimization of privacy protection, inference accuracy, and computational efficiency under a 128-bit security level.