Privacy Partition: A Privacy-Preserving Framework for Deep Neural Networks in Edge Networks

Abstract

The rise of the Internet of Things (IoT) encourages an emerging computing paradigm - edge computing - which leverages innovations in "last mile" communications infrastructure to provide improved quality of service guarantees to compute-intensive services such as autonomous driving and improved support for connected devices. Many high-value edge computing applications benefit from an integration of privacy-sensitive resource-constrained local data streams and data-hungry resource-constrained analytic tools like deep neural networks. We propose a practical method for privacy-preservation in deep learning classification tasks based on bipartite topology threat modeling and an interactive adversarial deep network construction in the context of edge computing. We term this approach Privacy Partition. A bipartite topology consisting of a trusted local partition and untrusted remote partition provides an apt alternative to centralized and federated collaborative deep learning frameworks in the case of deployment contexts such as IoT smart spaces, where users would like to restrict access to high-resolution data streams due to privacy concerns but would still like to benefit from deep learning services and external computational resources such as remote cloud data centers.