With the emergence of smart cities, Internet of Things (IoT) devices as well as deep learning technologies have witnessed an increasing adoption. To support the requirements of such paradigm in terms of memory and computation, joint and real-time deep co-inference framework with IoT synergy was introduced. However, the distribution of Deep Neural Networks (DNN) has drawn attention to the privacy protection of sensitive data. In this context, various threats have been presented, including black-box attacks, where a malicious participant can accurately recover an arbitrary input fed into his device. In this paper, we introduce a methodology aiming to secure the sensitive data through re-thinking the distribution strategy, without adding any computation overhead. First, we examine the characteristics of the model structure that make it susceptible to privacy threats. We found that the more we divide the model feature maps into a high number of devices, the better we hide proprieties of the original image. We formulate such a methodology, namely DistPrivacy, as an optimization problem, where we establish a trade-off between the latency of co-inference, the privacy level of the data, and the limited-resources of IoT participants. Due to the NP-hardness of the problem, we introduce an online heuristic that supports heterogeneous IoT devices as well as multiple DNNs and datasets, making the pervasive system a general-purpose platform for privacy-aware and low decision-latency applications.
Download