Privacy-Preserving Compressive Model for Enhanced Deep-Learning-Based Service Provision System in Edge Computing

Abstract

Advancements in the Internet of Things promote edge computing, which provides enhanced IoT services. IoT applications remarkably benefit from deep learning (DL) based on neural networks. Considering edge nodes with the limited ability of processing, the small-scale compressive DL model is applicatively deployed into them. However, the model can expose the privacy of training dataset. Thus, the model should be regarded as confidential due to the sensitive training dataset when applied in commercial and security fields. In this paper, we design a deep-learning-based service provision system for protecting the privacy and enhancing services in edge computing. We propose a practical approach for building the private compressive DL model by incorporating differential privacy mechanism on the cloud server side. For protecting the privacy of training dataset, the approach is composed of the private dense training step and the private compressive training step, where differential privacy is used to construct the private pre-trained dense model and the private compressive model, respectively. The private compressive model with 1/9 of the dense model's size can be practically embedded into edge servers. Edge servers provide enhanced services to the near IoT devices. Finally, we mainly execute a set of experiments on MNIST dataset. Compared to the dense model with no privacy constraints and with differential privacy (train and test accuracies of 98.0% and 98.2%, 95.5%, and 97.2%, respectively), the private compressive model perturbation with (6, 10-5)-differential privacy (train and test accuracies of 96.3% and 95.8%) can achieve high utility while upholding tight privacy.

Keywords

• Compressive model
• differential privacy
• deep learning
• edge computing