Dynamic Workload Selection for Ranked Fault Injection via Sequential Bernoulli Discovery Modeling

Fault Injection (FI) is the standard methodology for assessing neural-network resilience to hardware faults, but its cost remains high because each injected fault must be exercised over many input samples. Recent workload-ranking approaches reduce this cost by identifying the smallest set of input samples that maximizes fault detection. Despite their efficiency, these techniques remain static in terms of workload size. However, it has been noticed that the number of unique detected faults decreases as long as we apply input samples, leading to the possibility of dynamically adjusting the workload size. This paper proposes the dynamic adaptation of workloads by modeling the FI campaign as a sequential Bernoulli discovery process. For each processed input sample, a binary event indicates whether at least one previously unseen fault has been detected. Dynamic estimation of such discovery probability is then used as an early stop criterion for the FI campaign. Experiments on five network–dataset pairs, including CIFAR-10, CIFAR-100, MNIST, ImageNet, and a quantized CNN, show that dynamic estimation saves 77.8% of the workload while preserving 100% mean fault coverage.