Abstract: In order to overcome the shortcomings (e.g. low recognition accuracy and poor localization ability) of existing video forensics techniques, a two-stage algorithm is proposed based on Inception-V3 network, whose advantage lies in the fact that it can accurately identify a forged video and locate its forged frames. After an extensive research, it is found that the average value of all the pixels in a pristine video sequence is always larger than its forged one after the video sequence is processed via several operations such as high-pass filtering and convolution. To this end, in the first stage, a simple algorithm is proposed in which a predefined threshold is employed to distinguish a forged video and a pristine video. Considering the fact that features of each frame need to be extracted manually in existing algorithms, the dense convolution kernel structure of Inception-V3 network is adopted in the second stage to automatically extract high dimensional and multi-scale features of each forged frame. Inception-V3 network can accurately locate forged frames in a forged video since dimensional and multi-scale features can more adequately express the input information. The experiments show that the proposed method performs very well both in forged video identification and forged frame localization.

Key words: digital video forensics, video forgery detection, Inception-V3 network, forged frame localization