Network video streaming such as video hosting websites and connected cameras are increasingly popular. These applications bring new challenges to network administrators, such as leakage of privacy information and exposure to inappropriate content. When network administrators do not have control over end-hosts or processing video at individual end-hosts results in duplicate computation or a large configuration burden, they can mitigate such risks by inspecting video stream on the wire through Deep Packet Inspection (DPI). While the rise of Software-Defined Networking (SDN) and middleboxes brings new functionalities to existing networks, they are usually limited to fixed field processing.

In this thesis, we summarize the performance, architectural and format characteristics of network video streaming, design and implement the MiBao Video Processing Middlebox, and evaluate its overhead, effectiveness and scalability. Through allowing MiBao to vary the parameters of a face detection algorithm with respect to load and leveraging on GPU acceleration, MiBao can perform on-the-wire face detection and blur for up to four 1080p video streams at 25 FPS with less than 100ms playback delay using an AWS g2.2xlarge EC2 instance. As a proof-of-concept, MiBao shows that with GPU and tunable algorithms, middleboxes can meet video packets' delivery deadlines while performing complex computations.




Download Full History