Cosen: Efficient Collaborative Sensing with Heterogeneous Neighboring IoT Devices

Xingyu Feng, Chengwen Luo

Abstract: The increasing demand for edge computing has led to the deployment of artificial intelligence (AI) models on IoT devices. However, due to data privacy and communication cost concerns, AI models are often offloaded from the cloud to the edge or device side. To improve the sensing capability of local models, this paper proposes a collaborative inference framework, Cosen, which organizes collaborative inference groups among multiple neighboring IoT devices. The framework utilizes group decisions to achieve higher accuracy and more robust global inference. Focusing on improving the efficiency of collaborative inference, this paper proposes a model design scheme based on inference time constraints to design local models for heterogeneous IoT devices. A model training method based on knowledge distillation is used to improve the accuracy of the local model. Cosen also proposes a dynamic deep ensemble management approach that further improves the robustness of the system. Preliminary experimental results show that Cosen achieves higher accuracy and robustness for global inference.