Design of a machine learning layer to optimize forgery detection in videos

Abstract

With the advent of so many YouTubers, the number of videos uploaded to the cloud is increasing at an exponential rate. Most of the video creators take advantage of the fact, that the videos uploaded are not checked for any kind of tampering, which is due to the computational and information limit over the cloud. Due to this, there is lot of fake-like videos over the internet, which misleads the viewers into believing whatever the video creator conveys. All this leads to a lot of chaos and misunderstanding for the viewers at large. To overcome this issue, this text proposes a novel machine learning layer which is computationally efficient and uses a two-layer processing scheme for forgery detection in large scale videos. With the help of this scheme, the system pre-learns the forgery types, and keeps a large learning set ready for speedy reference, this initially static set, is then dynamically modified based on the real time videos given to the system. This two-layer scheme allows for an improved end-to-end delay for the system, without compromising on the system security. Our evaluations on the REWIND dataset show that the proposed system gives more than 95% accuracy in terms of forgery detection, and reduces the delay by more than 50% when compared to other state-of-the art algorithms like key point selection, genetic algorithm and others.