Title:Secure Aerial Surveillance using Split Learning

Abstract:Personal monitoring devices such as cyclist helmet cameras to record accidents or dash cams to catch collisions have proliferated, with more companies producing smaller and compact recording gadgets. As these devices are becoming a part of citizens' everyday arsenal, concerns over the residents' privacy are progressing. Therefore, this paper presents SASSL, a secure aerial surveillance drone using split learning to classify whether there is a presence of a fire on the streets. This innovative split learning method transfers CCTV footage captured with a drone to a nearby server to run a deep neural network to detect a fire's presence in real-time without exposing the original data. We devise a scenario where surveillance UAVs roam around the suburb, recording any unnatural behavior. The UAV can process the recordings through its on-mobile deep neural network system or transfer the information to a server. Due to the resource limitations of mobile UAVs, the UAV does not have the capacity to run an entire deep neural network on its own. This is where the split learning method comes in handy. The UAV runs the deep neural network only up to the first hidden layer and sends only the feature map to the cloud server, where the rest of the deep neural network is processed. By ensuring that the learning process is divided between the UAV and the server, the privacy of raw data is secured while the UAV does not overexert its minimal resources.