Privacy-Preserving Video Classification with Convolutional Neural Networks

Abstract

Video classification using deep learning is widely used in many applications such as facial recognition, gesture analysis, activity recognition, emotion analysis, etc. Many applications capture user videos for classification purposes. This raises concerns regarding privacy and potential misuse of videos. As videos are available on the internet or with the service providers, it is possible to misuse them such as to generate fake videos or to mine information from the videos that goes beyond the professed scope of the original application or service. The service provider on the other hand typically cannot provide the trained video classification model to be run on the client's side either, due to resource constraints, proprietary concerns, and the risk for adversarial attacks. There is a need for technology to perform video classification in a privacy-preserving manner, i.e. such that the client does not have to share their videos with anyone without encryption, and the service provider does not have to show their model parameters in plaintext. In this thesis, we propose privacy-preserving single frame-based video classification with a pre-trained convolutional neural network based on Secure MultiParty Computation. The pipeline for securely classifying a video involves three major protocols: oblivious selection of frames in a video, securely classifying individual frames in the video using existing protocols for image classification, and secure label aggregation across frames to obtain a single class label for the video. We perform run-time experiments for the use case of emotion detection in a video to demonstrate the feasibility of our proposed methods in practice.