PrivStream: Enabling Privacy-Preserving Inferences on IoT Data Stream at the Edge

Differential privacy (DP) has been recognized as a useful technique to achieve the trade-off between privacy guarantee and data utility. However, applying DP on the continuous data stream to protect sensitive inferred information faces two significant challenges. Firstly, directly adding noise into the continuous data stream may significantly reduce the data utility. Secondly, it is difficult to effectively and efficiently sample the data stream since the distribution of the data stream is usually unknown in advance. In this paper, we propose PrivStream, a privacy-preserving IoT data stream analytical framework based on edge computing, to address the two challenges. Specifically, PrivStream is split into two parts distributed at the IoT device and edge server, respectively. The first part, deployed at the device side, adaptively samples the data stream, leverages a tailored autoencoder to conduct data minimization and obfuscates the learnt features by injecting Laplace noise. The second part, deployed at the edge server side, reconstructs the perturbed features into obfuscated data stream for useful inferences without privacy disclosure. Moreover, we present the implementation details of PrivStream on a practical IoT system and theoretically analyze that Privstream can provide provable privacy guarantee. The experimental results based on realistic IoT data stream demonstrate that Privstream can preserve the data utility of useful inferences and mitigate the sensitive inferences simultaneously with efficient system overheads.