WESCO: Weight-encoded Reliability and Security Co-design for In-memory Computing Systems

Non-volatile memory (NVM) based in-memory computing (1MC) systems can avoid expensive data movement by implementing matrix-vector-multiplication calculations in memory, significantly reducing the power consumption and memory bandwidth requirements of deep neural networks (DNNs). Due to the non-volatility and the limited endurance of NVM devices, the system is ideal for low-power and retrain-free applications. However, NVM devices have reliability problems caused by device faults and data security risks due to non-volatility, making the system unreliable and unsecure. We observe that the impact of high-bit faults (HBFs) of quantized weights is far greater than low-bit faults (LBFs) on the classification accuracy of DNNs. Leveraging this observation, this paper proposes a lightweight and efficient co-design of reliability and security for retrain-free IMC systems, called WESCO, that can simultaneously tolerate faults and obfuscate the network. The weight matrices are encoded in row level by swapping the HBFs into LBFs to reduce the impact of faults on network accuracy without retraining; meanwhile, the implementation of our HBF and LBF swapping simultaneously obfuscates the network, so that the models cannot be accurately extracted from the stolen weights. The experimental results demonstrate WESCO can restore the classification accuracy of the DNN models to the baseline level at high fault rate of 5E-3 with a low area overhead of 1.17%, and limit the possibility of attackers stealing the model to infeasible brute force attacks.