Motion detection side-channel attacks on encrypted real-time video

Nimrod Harris, Yuval Elovici, Niv Gilboa

Ben-Gurion University of the Negev, 2018

We study information leakage from the encrypted video stream of surveillance cameras. The leakage is due to the correlation between the magnitude of motion that the camera captures and both its video compression and real-time transmission. We train classifiers on the encrypted video stream of surveillance cameras to achieve two tasks: estimating the fraction of the video frame in which motion occurs and estimating the number of people walking in a room that the camera monitors. For the first task we divide the frame into n equal parts and require the classifiers to identify how many of these parts include motion. The classifiers are accurate 99 percent of the time when n= 4 and 85 percent of the time when n= 16, as long as the camera does not reach its relatively low limit on bandwidth which reduces both video quality and information leakage. For the second task the classifiers are accurate 80− 90 percent of the …