A Tool for Privacy-Aware Online Personal Photo Sharing Using Deep Learning Technique

摘要

In recent years, online social networking has been considered as a sharing information platform and has occupied an essential part in many individual's lives and business growths. Consequently, there has been a marked increase in the collection and illegal exploitation of photos amassed online without owner consent, thus violating individual privacy rights such as contravention of online published photo laws which contribute to public social anxiety. To address these concerns, we propose a face recognition tool based on the Deep Learning Convolutional Neural Network (CNN) technique, which may be utilized within a social networking website as a gateway control for posting images. The goal of this paper is to preserve user privacy by preventing their images from being posted on social networking sites without prior consent. This tool will extract features from an input photo posted on a social network site and compare those attributes against the facial characteristics of photos in a prohibited dataset, which is comprised of users unwilling to share their photos. Depending on the result, the CNN-based tool could either allow sharing of the photo or prevent and alert the user attempting to post or share a given photo about his/her potential violation of end-user privacy provided the image belongs to a person on the banned list. Additionally, the CNN tool will provide an option for a user to add his/her photo to the banned list. The proposed tool includes two main elements which have been developed in Python with Jupyter Notebook. The first component is a deep learning model which is trained on LFW images dataset capable of achieving 91.89% matching accuracy. The second is the GUI of the tool which allows the user to input photos and use the trained model to predict whether this photo belongs to the person in banned list, thus preventing illicit sharing downstream. The integration between two elements has been tested and achieved 85% accuracy.