Functional Self-Test for Deep Neural Networks

As deep neural networks (DNNs) are increasingly deployed in safety-critical systems, ensuring their reliability under in-field conditions is essential. Although inference engines such as GPUs and TPUs undergo structural testing, hardware faults escaping from such tests can still degrade the performance of deployed models in the field. In this work, we propose Neural Built-In Self-Test (NBIST), a small neural network generating on-the-fly test patterns for the Network Under Test (NUT) on the same inference engine. Unlike traditional structural hardware BIST, NBIST performs functional self-testing by leveraging a co-deployed test generator that runs entirely in software on the inference hardware, requiring no dedicated hardware test circuitry. NBIST maps pseudo random numbers to structured test patterns that achieve up to 100% fault coverage while, aside from storing intended labels, requiring only a constant memory overhead for storing generator weights and seed. NBIST enables in-field and on-demand testing without relying on pre-stored ATPG patterns. Experiments across multiple models, tasks, and fault scenarios show that NBIST consistently outperforms both random and adversarial tests in terms of coverage (up to 100%) and storage efficiency compared to storing test patterns directly.