Efficient and privacy-preserving deep inference towards cloud–edge collaborative

Abstract

The cloud–edge collaborative inference approach splits deep neural networks (DNNs) into two parts to run collaboratively on resource-constrained edge devices(AIoT devices) and cloud servers, aiming at minimizing inference latency and protecting data privacy for AIoT computing system. However, despite not exposing the raw input data from edge devices directly to the cloud, state-of-the-art attacks can still target collaborative inference to reconstruct the raw private data from exposed local models’ intermediate outputs, introducing serious privacy risks. In this paper, we propose a secure privacy inference framework for cloud–edge collaboration system towards AIoT network, called CIS (Collaborative Inference Shield), which supports adaptively partitioning the network according to dynamically changing network bandwidth and fully releases the computational power of edge devices. To mitigate the influence introduced by private perturbation, CIS provides a way to achieve differential privacy protection by adding refined noise to the intermediate layer feature maps offloaded to the cloud. Meanwhile, given a total privacy budget, the budget is reasonably allocated by the size of the feature graph rank generated by different convolution filters, making cloud inference robust to the perturbed data, thus effectively trading-off between privacy and availability. Finally, we construct a real cloud–edge collaborative inference computing scenario to verify the effectiveness of inference latency and model partitioning on resource-constrained edge devices. Furthermore, the state-of-the-art cloud–edge collaborative reconstruction attack is utilized to evaluate the practical availability of the end-to-end privacy protection mechanism provided by CIS.