Efficient Software-Implemented HW Fault Tolerance for TinyML Inference in Safety-critical Applications

TinyML research has mainly focused on optimizing neural network inference in terms of latency, code-size and energy-use for efficient execution on low-power micro-controller units (MCUs). However, distinctive design challenges emerge in safety-critical applications, for example in small unmanned autonomous vehicles such as drones, due to the susceptibility of off-the-shelf MCU devices to soft-errors. We propose three new techniques to protect TinyML inference against random soft errors with the target to reduce run-time overhead: one for protecting fully-connected layers; one adaptation of existing algorithmic fault tolerance techniques to depth-wise convolutions; and an efficient technique to protect the so-called epilogues within TinyML layers. Integrating these layer-wise methods, we derive a full-inference hardening solution for TinyML that achieves run-time efficient soft-error resilience. We evaluate our proposed solution on MLPerf-Tiny benchmarks. Our experimental results show that competitive resilience can be achieved compared with currently available methods, while reducing run-time overheads by similar to 120% for one fully-connected neural network (NN); similar to 20% for the two CNNs with depth-wise convolutions; and similar to 2% for standard CNN. Additionally, we propose selective hardening which reduces the incurred run-time overhead further by similar to 2x for the studied CNNs by focusing exclusively on avoiding mispredictions.