A Data-Driven Approach to Online Fault Detection in RRAM-Based Neuromorphic Hardware Using Adversarial-Inspired Test

Resistive random access memory (RRAM)-based neuromorphic computing systems (RNCS) show great promise for low-power edge computing applications. However, hardware reliability issues in RRAM devices remain unresolved. Due to the inherent fault tolerance of RNCS, many faults do not immediately degrade system accuracy, but minor faults may pose latent risks. Meanwhile, traditional fault detection methods are insufficient for edge systems that require strict efficiency and real-time operation. This work proposes a data-driven approach based on multistage extreme gradient boosting (XGBoost) and adversarial-inspired test, specifically designed for fast online fault detection in RNCS, focusing on cell-level SAF detection under diverse fault distributions. Experiments on convolutional layers of pretrained convolutional neural network (CNN) models in RNCS demonstrate the framework’s effectiveness for real-time SAF diagnosis. Under low fault rate conditions, the proposed approach achieves an average recall of 99.8% and an average$F1$-score of 99.5%. In addition, the test dataset generated using the iterative fast gradient sign method (IFGSM) (adversarial-inspired test) effectively improves the distinguishability of RNCS fault-induced output signatures.