Edge AI Trustworthiness: Revisiting Bit-Flip Impacts and Criticality Conditions

Neural networks (NNs) are increasingly used in embedded systems driven by artificial intelligence (AI) features. Such systems are now developed for many critical applications with strong dependability constraints and trustworthiness requirements. The effect of bit-flips occurring during inferences in the field has therefore become a major concern. Dedicated design methods have been proposed to increase the robustness of NNs and enforce trustworthiness, while minimizing implementation overheads in the context of Edge AI. Such approaches are usually based on criticality criteria characterizing the most impactful bit-flips. In this work, the two main criteria are revisited in the case of bit-flips in data that is independent of the hardware/software implementation. Extensive statistical fault injection results are analyzed for a case study based on a quantized version of LeNet. They first demonstrate that any criterion is only held in some perturbation conditions, depending on the type of perturbed data and on the perturbed layer. The data representation format has also a significant impact. Surprising outcomes of the case study allow identifying a new parameter overlooked in the literature, i.e., the definition of the activation functions. Another important finding, mostly neglected in the literature, is the large number of bit-flips having a positive impact on inferences by correcting misclassifications of the nominal NN. In the presented case study, almost 40% of the misclassifications due to bit-flips are compensated for by the positive effects of other bit-flips, leading to only 0.06% global accuracy loss when bit-flips occur. These outcomes indicate that mitigations must be carefully tailored to each layer and data subset of the NN in order to accurately limit the effect of critical bit-flips. Moreover, at the same time they must avoid suppressing the advantage taken from bit-flips with positive effects. Although derived from a specific case study, these findings have a wide significance in design practice.