Deep Learning based Effective Identification of EU-GDPR Compliant Privacy Safeguards in Surveillance Videos

With video surveillance through Closed Circuit Television (CCTV) cameras becoming increasingly ubiquitous, maintaining the integrity of such videos assumes significance to guarantee compliance with privacy protection laws. Recently adopted European General Data Protection Regulation (EU-GDPR) suggests encryption as a reversible safeguard for all type of data protection. Traditionally, data encryption methods deal with naïve encryption of the videos which is not advisable if the intention is to provide efficient privacy to some important parts/areas in the video. This has resulted in the research being directed towards Features-of-Interest (FOI) based privacy protection schemes, wherein only selected parts of a video frame are encrypted/obfuscated. This research utilizes a deep learning (DL) model to identify GDPR compliant privacy-preserved (encrypted) videos among different types of obfuscated videos. For evaluation, two types of video obfuscation methods viz. selective encryption and data removal are applied on the extracted background and foreground FOIs/parts of video frames. Then the applicability of a modified version of the highly-optimized MobileNetV2 model for identification of privacy-preserving features is presented. Results show that the lightweight DL model is suitable for deployment on low-resource hardware (such as Raspberry Pi's) in Internet of Things (IoT) infrastructure in smart cities.