An improved surveillance video forgery detection technique using sensor pattern noise and correlation of noise residues

Abstract

Digital videos have become an important aspect of our lives lately, from a personal memorable to surveillance videos which can be presented in a court as an evidence now, this video evidence can be very important for the court of law and the investigators to understand the events as they occurred. Huge inventions in the video editing world have resulted in a margin of doubt, to the point where courts are now reluctant to accept any video evidence in to their case files, as for a digital video to be presented as an evidence, it’s integrity and authenticity is very important. This progression in technology can push the courts to a point where they refuse to trust anything that was caught on a video. Thus, more efforts have to be made in the video forensics world to avoid such a future. It has been shown that sources induce their specific noise patterns (i.e. Sensor pattern noise) to the captured media, and these patterns can be used to trace the source sensor and also to detect any forgeries. Acquiring SPN (sensor pattern noise) consists of two steps, first to get noise residues by subtracting denoised frames from the actual frames taken from the sample data and then to average those noise residues to get the sensor pattern noise. Problem arises when the attacker can also extract sensor pattern noise or has access to it. The attacker can induce SPN in forged frames which will prevent detection using the previous SPN based detection technique. This paper first proposes how the attacker can compensate a forged image by inducing SPN in it and then proposes a forgery detection technique for such a scenario which would not only rely on noise residue correlation with SPN but correlation with noise residue from the previous frame as well. We first estimate sensor pattern noise based on locally adaptive discrete cosine transform, then we correlate noise residue of frame n under investigation with the sensor pattern noise and with the noise residue of the previous frame n-1 as well. Now, even if the attacker compensates the forged frame by inducing SPN in it, the inter-frame continuity of noise shall be disturbed hence identified. Simulations using benchmark dataset for SPN extraction technique and custom dataset for video forgery detection, show that this methodology successfully detects any forgeries and locate their positions in the normal case and in the case where SPN is induced in the forged frames as well.