Reliability Estimation of Split DNN Models for Distributed Computing in IoT Systems

Split Computing (SC) is an attractive solution toward achieving an optimal balance between onboard capabilities, resource usage, and performance in Deep Neural Networks (DNNs) in the context of Internet-of-Things applications. However, the reliability of SC-based systems has been completely overlooked in the literature until now, especially when considering hardware faults arising during the system operational lifetime (e.g., due to aging). This work presents a methodology to estimate the reliability of different SC configurations produced in the DNN design phase, enabling the exploration of trade-off between dependability and typical performance metrics. Our approach is analytical, and takes into account the architecture of the DNN and the possible reliability interdependencies it induces. Specifically, we derive a general probabilistic equation based on the building blocks (and their interconnections) of the DNN and the sensitivity to hardware faults of each layer. In the SC setting, a probabilistic equation can be derived considering different fault rates for each hardware component of the system (i.e., mobile device and server). Based on the proposed approach, we evaluate splitting options for a computer vision task based on the multi-dimensional trade-off between energy consumption, channel usage, and reliability.