SCIoT: Design and Evaluation of a Split Computing Framework for Collaborative Inference in the IoT

The increasing reliance on machine learning in Internet of Things systems demands to evaluate the trade off between computing on the resource constrained devices or offload the computation to more powerful edge devices. Split computing has emerged as a promising paradigm to bridge this gap by partitioning workloads between resource-constrained devices and edge devices in a flexible way. Existing approaches, however, often remain tied to specific model architectures or assume theoretical network and device conditions, hence limiting their applicability in realistic deployments. In this paper, we introduce SCIoT, a framework for Split Computing in the Internet of Things that seeks to address these shortcomings. SCIoT enables flexible and adaptive partitioning across heterogeneous devices, explicitly accounting for resource availability, fluctuating network performance, and data sensitivity. The framework incorporates dynamic policies that balance latency, bandwidth usage, and privacy, moving beyond static or one-size-fits-all strategies. We evaluate SCIoT across representative scenarios, demonstrating its ability to adaptively reconfigure computations while maintaining competitive efficiency. Our results show both the advantages and the current limitations of split computing in practice, contributing a step toward more robust, adaptive, and privacy-aware collaborative inference in IoT ecosystems.