Joint Model, Task Partitioning and Privacy Preserving Adaptation for Edge DNN Inference
Jingran Jiang, Hongjia Li, Liming Wang · 2022 IEEE Wireless Communications and Networking Conference (WCNC) · 2022
Deep Neural Networks (DNNs) have been widely used in everyday life owing to their impressive performance in complex machine learning tasks. The performance however comes at the cost of high computational complexity, which hinders the application of many DNN models in resource-constrained Internet-of-Things (IoT) and mobile devices. Device-edge collaborative DNN inference (referred to as co-inference) is an effective way to address the issue. However, it requires non-trivial algorithmic design, since the compound performance indicators including the inference efficiency and accuracy, and the data privacy have to be jointly considered. In this paper, we extend the degree of flexibility of the classical co-inference schemes, and propose a joint model, partitioning point and privacy differential intensity adaptation framework for co-inference, which comprises of the offline and online phases. In the offline phase, we train the co-inference model set that consists of a series of sub-models with different complexities, and profile the necessary performance of the sub-models. On that basis, we design an efficient algorithm for the online phase to promptly choose the sub-model, partitioning point and privacy differential intensity to meet the latency constraint and achieve the optimal accuracy-privacy tradeoff. Finally, extensive evaluations are carried out to demonstrate the effectiveness of our proposed framework.