Assessing the Reliability of Different Split Computing Neural Network Applications

Neural Networks (NNs) are increasingly adopted in many domains. Given their considerable computational expenses required only during inference, it has been proposed that their architecture can be split into two sections (Head and Tail) in a paradigm referred to as Split Computing (SC). The Head section executes the first part of a NN at a mobile device, while a Tail model calculates the missing part of the NN resorting to a server. SC allows a designer to find the optimal configuration, trading off computing load, performance, latency, and power. However, the adoption of SC approaches in safety-critical domains involves strong reliability concerns. This paper introduces an environment allowing to estimate the effects of permanent faults affecting the hardware supporting the execution of the head model, which is the most prone to faults. Results are reported for different NNs and different configurations, demonstrating the effectiveness of the method and paving the way towards the deployment of more resilient SC systems. Our results show that the considered split configurations of all models under test are up to 10% more critically vulnerable to faults than the corresponding original architecture.