Video forensics: a study of multimedia forensics with emphasis on modern video

Brian C. Hosler

doi:10.17918/00010988

Multimedia content is more accessible than ever before due to a variety of factors including more advanced compression techniques, higher throughput channels, and an increase in media sharing hardware and software. The ease with which people can send and receive media is associated with an increase in the amount of false or misleading media being peddled by malicious actors as fact. It is in our interest to discriminate between pristine and manipulated media and, by proxy, real and fake media. In particular, the rapid proliferation of video content has outpaced out ability to manually verify the truth represented in such content. The field of multimedia forensics is dedicated, in part, to developing traces, tools, and techniques for determining media provenance, authenticity, and originality. In this thesis we propose new techniques for multimedia forensics with an emphasis on the growing medium of video. The first technique we propose is a method of performing video source camera model identification. The video capture and compression pipeline native to each video camera model cane be modeled as a series of signal processing operations which are known to leave forensic traces in the resultant media. The metadata of this media can be falsified so it is necessary for a forensic analyst to determine the source camera model from the pixel values themselves. To do this we develop a machine learning based algorithm that fuses multiple individual forensic traces, leveraging the large nature of video to increase accuracy. We discuss a series of experiments performed to validate the efficacy of our approach and the demonstrate importance of fusion. Our second approach addresses the problem of identifying generated content. Specifically, we consider content that was generated using deep-learning-based video and audio generation. We propose to detect this fake content through analysis of the temporal intra-medium and inter-medium content. Specifically, each medium, audio and video, is projected into a lower dimensional space representing the instantaneous emotion over time. We show that the evolution of these emotions over time, as well as the disparity in emotion between mediums can be used to identify generated content. While deep-learning based techniques for video forgery are becoming increasingly common, video forgery through traditional video editing is easier to produce and similarly effective. In our next approach we propose a method for detecting video speed manipulation performed on an encoded video. Video compression relies on redundant video frame information; as frames are added or removed to perform speed manipulation, the effectiveness of the compressor is likely to change as well. In this approach we model the effect speed manipulation has on the number of bytes used to encode a frame, and conduct experiments to show that this model can be used to perform speed manipulation detection and even estimate the parameter of manipulation. While effective, this approach is predicated on the video structure and frame encoding order. In our next approach we discuss this limitation, and perform experiments to determine its importance. We improve upon the model of the initial approach and increase its performance. This initial approach is also subject to failure via anti-forensic attack. In this approach we create such an attack, and propose an anti-anti-forensic defence of such an attack. Finally, we propose a new tool for extracting forensic traces from an encoded video. Our previously proposed techniques leveraged information from videos which is not typically available to the viewer. We discuss the video encoding process and highlight elements of the process which we theorize to have significant forensic applications. We conduct experiments with this tool to demonstrate our hypothesized traces and how they are affected by video editing operations.

Video forensics: a study of multimedia forensics with emphasis on modern video

Files and links (1)

Abstract

Metrics

Details

Video forensics: a study of multimedia forensics with emphasis on modern video

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media