M-EL2N: Data-Efficient Fault Injection Through Sample Difficulty Analysis

Fault Injection (FI) is the standard technique for assessing the resilience of Convolutional Neural Networks (CNNs), but campaign time scales with the number of input stimuli applied per injected fault. Prior work has shown that selecting challenging inputs based on confidence metrics can significantly reduce this workload. In this paper, we propose Mirrored EL2N (M-EL2N), a novel sample-ranking metric that symmetrizes the original EL2N score by re-scaling the contribution of misclassified samples. This yields a unified notion of sample difficulty that prioritizes highly informative inputs regardless of prediction correctness, enabling faster exposure of critical faults, and requires no architectural knowledge. We evaluate Mirrored EL2N against established metrics (Cross-Entropy, DeepGini, EL2N) across multiple CNN architectures and datasets under a single-bit fault model in network weights, and we present initial results on a post-training quantized CNN. Experimental results show that Mirrored EL2N consistently achieves earlier detection of critical faults with up to three orders of magnitude reduction in the number of test samples required to reach 99% of the detected critical faults, while maintaining near-perfect detection curves (AUC ≈ 1). These results establish Mirrored EL2N as a practical and lightweight tool for data-efficient FI workload reduction in both floating-point and quantized CNNs.