Enhancing the Reliability of Split Computing Deep Neural Networks

Artificial intelligence is becoming increasingly popular for IoT applications in safety-critical fields (e.g., autonomous systems and biomedical, robots). Unfortunately, the inference’s workload process alone increases as the model size grows. To meet the computational power limitations of mobile devices running IoT applications, modern services sometimes resort to the Split Computing paradigm. Split Computing divides the inference process of a Neural Network into Head and Tail for their execution in a mobile device and a server, respectively, which also allows the reduction of the overall IoT device’s computational cost. Nonetheless, Split Computing can be used in safety-critical fields where reliability is crucial, especially when mobile devices have computational and cost restrictions. This paper introduces hardening techniques acting on the software to mitigate the effects of hardware faults on Split Computing models. The proposed hardening techniques consist of i) a bounded activation function whose thresholds are refined by training, and ii) a per-channel bounding of the bottleneck quantization of the split points. To quantitatively assess their effectiveness, we resorted to two different split configurations of a model for image classification. In addition, we considered a Split Computing model for object detection. Our findings indicate that the proposed approaches effectively reduces fault effects by 3.5% for image classifiers and 5.73% for object detectors when compared with other hardening approaches for general DNNs.