Obfuscated image classification for secure image-centric friend recommendation

Abstract

Image sharing is one of the most popular activities of smart citizens in recent time. People tend to frequently upload images through smart devices to express different aspects of their life with connected peers using the image sharing feature facilitated by different social media services such as Facebook, Flickr, Pinterest, and Instagram, etc. The service providers of such web services are utilizing the image features and high availability of cloud storage at low cost for providing friend recommendation service to build online smart community. However, after reports of citizens surveillance by government agencies and the celebrity photo leakage incident in iCloud, users have become concerned about their photo privacy. In addition to that, the cloud services are vulnerable to security threats and an adversary capable of breaching the security would be able to access the images residing in the cloud storage. Obfuscating images before sharing can be considered as a potential solution; however, de-obfuscating and then classifying millions of images at the service provider's end does not provide privacy guarantee and it is computationally expensive. Motivated by this scenario, we propose a practical privacy-preserving image-centric friend recommendation framework compatible with smart devices which protects the privacy of images through obfuscation and classifies obfuscated images using the deep neural network to build user profiles for friend recommendation.