Estimation and concealment of forensic multimedia signatures

Abstract

The remarkable evolution of digital imaging techniques, processing and sharing in the past decades has spurred the penetration of multimedia into our lives. Unaccountable and ubiquitous use of multimedia brings severe issues and challenges about its origin and veracity. For instance, due to the growth of social media and instant messaging applications, the circulation of tampered content has become an unavoidable reality. The waning credibility of digital content has also lead to unfavorable consequences in terms of political, economic and social issues. Therefore, to address the source of origin and processing history-related issues of multimedia content, the scientific community has focused its attention on digital multimedia forensics techniques. Two of our major contributions belong to this class of forensics algorithms. On the other hand, to counter forensics algorithms, a parallel area of adversarial signal processing has also gained a lot of momentum. These counter-forensics algorithms are designed to hide the fingerprints left behind by image processing operations, thereby degrading the performance of forensics detectors. Our third major contribution belongs to this class of counter forensics techniques.

In our first contribution, we address the problem of establishing the link between a given image to its source camera device. This problem is called Source Camera Identification. In order to achieve this, the photo response non-uniformity (PRNU) characteristic of the sensor is exploited. However, the existing techniques suffer from the problem of noise induced during in-camera image processing. This noise can suppress the PRNU leading to poor camera identification performance. To this end, we propose a novel algorithm for robust estimation of PRNU from probabilistically obtained raw data. Since not all cameras provide raw data as their output, we compute raw data from the JPEG output using a probabilistic color de-rendering procedure. The estimated raw data is modeled as a Poisson process, and Maximum Likelihood Estimation is used for PRNU estimation. We also extend the estimated PRNU for tampering detection. The extensive experimental analysis performed on thousands of patches from various cameras reveal state-of-the-art performance. We also demonstrate the robustness of estimated PRNU by accurate tampering localization.

In our second contribution, we analyse the counter forensic algorithm for Contrast Enhancement, which is a common post-processing step in image tampering. The existing algorithms only consider the spatial domain. In our work, we consider changes in both spatial and DCT domain into account and obtain enhanced images such that the statistical properties are similar to natural images. Unlike the conventional techniques which lead to artifacts that can be captured by forensics detectors, the proposed algorithm suppresses the detectable artefacts. In our experiments, we demonstrate a significant performance degradation for deep learning as well as steganalysis-DCT feature based detectors. We also compute the popular image quality assessment metrics and show that the proposed model generates better visual quality images compared to the existing counter forensics techniques.

Nowadays cameras are also used in personal and commercial vehicles which pose a different problem of linking a given video to the vehicle in which the camera was mounted. This is useful for various applications, for example, insurance companies can authenticate the origin of video before processing the claim. In a different scenario of illegitimate video upload on the web, the video can be traced back to the car it originated from. To this end, we state our third contribution, in which we introduce the new area of multimedia vehicle forensics. We propose an algorithm for linking a dash-cam video to a specific car. Inspired by human gait bio-metrics, we observe that the subtle motion pattern of every vehicle can serve as its unique signature. We extract motion blur from dash-cam videos, which encode the motion pattern of the car. Experimental results on hours of dash-cam videos of several cars show the effectiveness of our approach. We further investigate the adversarial process of forging the signature of the vehicle and propose a forensics method to detect such forgery.