EISFINN: On the Role of Efficient Importance Sampling in Fault Injection Campaigns for Neural Network Robustness Analysis

Deep Learning (DL) systems have proliferated in many applications, requiring specialized hardware accelerators. An efficient fault-injection methodology is needed for analyzing the resilience of advanced DL systems against different types of faults, which can lead to undetectable and unrecoverable errors. Typically, in a fault injection campaign, the faults are sampled from the random uniform space covering all the possible faults. However, this method is extremely inefficient for large Deep Neural Networks (DNNs), and existing solutions require apriori knowledge on the model, filtering out the search space. Therefore, we propose EISFINN, a novel methodology that employs user-selected neuron sensitivity algorithms to generate importance sampling-based fault-scenarios. Without any a-priori knowledge of the model-under-test, EISFINN provides an equivalent reduction of the search space as existing works, while allowing long simulations to cover all the possible faults, improving on existing model requirements. Our experiments show that the importance sampling provides up to $10 \times$ higher precision in selecting critical faults than the random uniform sampling, in less than 100 faults.