Learning the Optimal Path and DNN Partition for Collaborative Edge Inference

Recent advances in Deep Neural Networks (DNNs) have enabled a wide range of intelligent mobile applications, but their computational demands pose challenges for resource-constrained devices. Collaborative edge inference addresses this by partitioning a DNN inference task into several subtasks across multiple network nodes. However, most existing methods assume known network parameters or fixed processing paths. In this paper, we consider a more complex setting where network parameters are unknown and multiple paths are available. The goal is to learn both the optimal path and the DNN layer assignment along it, while accounting for security threats and path-switching costs. We first derive structural insights under full information to reduce the decision space. We then formulate the learning problem as an adversarial group linear bandit with switching costs, where rewards follow a hybrid stochastic-adversarial process. To solve this, we propose B-EXPUCB, a new algorithm that combines ideas from blocked EXP3 and LinUCB, and show that it achieves sublinear regret. Extensive simulations demonstrate that B-EXPUCB outperforms existing methods in collaborative edge inference.