HEHLKAPPE: Utilizing Deep Learning to Manipulate Surveillance Camera Footage in Real-Time
A. Aigner, R. Zeller - HEHLKAPPE: Utilizing Deep Learning to Manipulate Surveillance Camera Footage in Real-Time - Proceedings of the 14th International Conference on Availability, Reliability and Security, Canterbury, CA, United Kingdom, Vereinigtes Königreich von Großbritannien und Nordirland, 2019, pp. 8
Image analysis and manipulation have always been active topics, both in practice and academia. Driven by the progress in the field of deep learning, significant advances have been achieved in recent years. This causes even complex image manipulation and analysis tasks to be easily applicable by a wide audience. Combined with traditional attack methods, this results in new attack vectors. In this paper, we present HEHLKAPPE, an application to hide persons from real-time video streams while keeping other movement untouched. The application is fully automated and does not require any domain knowledge in deep learning, image manipulation or other related areas in order to use it. In addition, we present 2 attack vectors to access the video stream to enable the manipulation. Our evaluation shows that HEHLKAPPE works well with static camera positions, is able to adapt to background changes and therefore is suitable to deceive observers. The discussion of the results discovers potential for improvement by using even more sophisticated techniques. We are confident that these techniques will be applicable in real-time in the near future. Appropriate countermeasures to mitigate our attack include improving the state of IoT security and verifying the authenticity of each frame using a blockchain-linke structure.