A novel half-complementary cipher for in-situ secure computing-in-memory

The advent of the new AI era has led to an increase in the number of intelligent terminals in use and the generation of a vast amount of data, driving the development of novel AI accelerator techniques. However, in the pursuit of high speed and high efficiency, the rising privacy risks and security concerns pose an urgent need for highly applicable encryption solutions without heavily increasing hardware cost for AI accelerators. In this work, to enhance the security of computing-in-memory architecture, we propose a novel half-complementary cipher to provide an effective information shield and in situ decryption function without sacrificing hardware cost. With the integrated 4 Kb STT-MRAM system, the decryption and inference of speech command classification is experimentally demonstrated, validating the feasibility of the half-complementary cipher. Simulation results for larger neural networks also substantiate the effectiveness. Moreover, the 2T2R structure-based half-complementary cipher facilitates further scaling up of the computing-in-memory architecture and offers a promising avenue for secure and efficient edge AI, thanks to a ∼50% reduction in three key factors, including the number of memory cells, the maximum accumulative source line current, and the noise in its worst-case scenario, compared to the conventional XOR cipher.