Video Forensics in Temporal Domain using Machine Learning Techniques

In defence and military scenarios, Unmanned Aerial Vehicle (UAV) is used for surveillance missions. UAV's transmit live video to the base station. Temporal attacks may be carried out by the intruder during video transmission. These temporal attacks can be used to add/delete objects, individuals, etc. in the live transmission feed. This can cause the video information to misrepresent facts of the UAV transmission. Hence, it is needed to identify the fake video from the real ones. Compression techniques like MPEG, H.263, etc. are popularly used to compress videos. Attacker can either add/delete frames from videos to introduce/remove objects, individuals etc. from video. In order to perform attack on the video, the attacker has to uncompress the video and perform addition/deletion of frames. Once the attack is done, the attacker needs to recompress the frames to a video. Wang and Farid et. al. [1] proposed a method based on double compression technique to detect temporal fingerprints left in the video caused due to frame addition/deletion. Based on double MPEG compression, here we propose a video forensic technique using machine learning techniques to detect video forgery. In order to generate a unique feature vector to identify forged video, we analysed the effect of attacks on Prediction Error Sequence (PES) in various domains like Discrete Cosine Transform (DCT), Discrete Fourier Transform (DFT), Discrete Wavelet Transform (DWT) domain etc. A new PES feature γ is defined and extracted from DWT domain, which is proven robust training parameter for both Support Vector Machine (SVM) and ensemble based classifier. The trained SVM was tested for unknown videos to find video forgery. Experimental results show that our proposed video forensic is robust and efficient in detecting video forgery without any human intervention. Further the proposed system is simpler in design and implementation and also scalable for testing large number of videos.