Edge AI Dependability: Is Bit-flip Direction the Right Concern in Trustworthiness Reinforcement?

Edge AI is today used in core features for many systems in almost all application areas, including critical ones. User trust implies guarantees about the system reactions when facing perturbations such as errors occurring in data bits during the inference computations. Many design approaches can be used to increase the system robustness. Recent works on neural networks assume or demonstrate on case studies that bit-flips are more critical in one direction (0 to 1) or when pushing the values further from zero; they therefore propose to tune hardening mechanisms taking advantage of such properties. In our case study, we revisit these assumptions and show that they do not hold in all cases. Our comprehensive fault injection campaigns based on a large dataset demonstrate that (1) there is no absolute most critical bit-flip direction (2) the impact of a bit-flip on the data value may better take into account some design decisions but (3) contradictory results shown in the literature and in this work require taking into account other characteristics. These results give insights into requirements for efficient mitigation strategies.