A Post-Quantum Encryption Mechanism Based on Convolutional Neural Network Accelerator

For most of the edge-based system-on-chips (SoCs), the inference (e.g. a CNN accelerator) and the security subsystems are typically separately designed and interacted with each other through a data bus. After the convolutional layer is computed, the bus transfers the plaintext intermediate data to the encryption module for encryption. The encrypted data is stored in memory for future computing in the cloud. In order to avoid transmitting plaintext on the data bus, make the design lighter, and resist attacks on current encryption mechanisms by quantum computers, this paper proposes a new architecture that combines hardware design of convolutional layers and inverse number theory transforms (INTTs), allowing the intermediate data of the inference process to be stored directly as ciphertext. The paper presents the hardware implementation of two networks, LeNet-5 and MobileNet v1, and tests them on the MNIST dataset and CIFAR-10 dataset, achieving accuracy rates of 95% and 92%, respectively. Compared with latest implementation solutions in post-quantum encryption mechanisms, we achieve a 25% reduction in computation time for INTT operations.