Image Forensics: Towards Effective Techniques for Image Authentication in Compressed Domain

Abstract

Nowadays, most images are tampered with in a lossy compressed scenario that can spread fake information on social media. On the other hand, image forensics forms an important sub-domain in detection of such image tampering, and it has various applications: surveillance, copyright protection, journalism, biomedical imaging, etc. Towards this, the research reported in this thesis primarily resides in image forensic analysis in the compressed domain. Further, the thesis examines the issues of manipulation detection, parameter estimation, and manipulation operator chain classification, which are some of the most crucial forensic applications. Therefore, this work is driven by manipulation detection and further move on to parameter estimation, which is relatively less established. In addition, another unsolved issue that required investigation was operator chain classification in JPEG compressed images. The preliminary contribution of this thesis deals with the manipulation detection problem, which is split into sub-problems of manipulation detection and its parameter classification. Manipulation detection focuses on separating altered images from unaltered ones, while parameter classification focuses on further classifying by which factor the doctored image is altered. Towards this, conventional Bartlett and Welch power spectral density (PSD)s are utilized to characterise the compressed domain discrete cosine transform (DCT) histogram. The first approach, Bartlett PSD of the DCT histogram and one-dimensional (1D) convolutional neural network (CNN) are proposed for manipulation detection and its parameter estimation. In contrast, the second approach proposes a Welch periodogram for quality factor estimation of re-compressed images. Both techniques performed well in all experiments using DCT coefficient histograms retrieved from altered and unaltered compressed images, while the Barlett method slightly outperformed the other. When the doctored input images were noisy, however, the performance of the PSD degraded, highlighting the need to investigate more sophisticated approaches capable of mitigating compression noise. To overcome this, the second contribution made an attempt to propose a new approach to the manipulation detection problem by utilizing a CNN-based denoiser to extract the noise residuals. Initially, a CNN-based denoiser with ten convolutional layers without pooling layers is used to capture correlation patterns of neighbouring pixels, which could suppress the blocking artefacts (BAR) and enhance the clues separately. Further, these residuals are fed to the feature conceptual extraction and classification stage to detect manipulation by eliminating the blocking artefacts. In addition, these noise residuals are also utilized for estimating scaling factors, which gives promising results, particularly for detecting downscaling scenarios. However, the proposed deep CNN fails to detect multiple manipulations. Hence, a generalized manipulation detection scheme has to be investigated. The latter part of the thesis shifted towards exploring the multiple manipulation detection problem, which consisted of different manipulations (JPEG-manipulation-JPEG) in most real-time applications. This premise led to exploring two frameworks (MDRNet and MPeRNet) to detect and estimate multiple manipulations by extracting noise residuals using residual blocks. In both frameworks, noise residual extraction stage significantly extracts manipulation traces by expanding the front-end detector that can exploit noise residuals by suppressing the image content. In addition, MDRNet and MPeRNet effectively detect multiple manipulations and estimate the parameters with shortcut connections from noise residuals. However, DCT residuals are not utilized in the above frameworks, which impacts the performance of manipulation operator chain detection under compressed scenarios. The last part of the thesis mainly concentrated on under-explored manipulation operator chain classification problem, which entails recognizing complex manipulations like Gaussian blurring and resampling. According to inferences taken from prior methods, the feature analysis could be enhanced if a representation scheme for the manipulation under compression were investigated rather than the uncompressed scenario. To achieve this, two approaches with multi-streams, such as manipulation-based and compression-based feature maps, are used to describe the tampered information of the doctored images. In the first approach, the spatial and noise streams are concatenated to obtain spatial features, and 25 DCT stream residuals pass through residual blocks to get frequency domain features. In the other approach, two streams, noise residual extraction (NRE) and compression feature extraction (CFE), are utilized to extract features of a doctored image. Finally, these feature maps are concatenated in both approaches and given to the classification stage to identify the operator chains. Both approaches are proposed for JPEG-resistant image operator chain detection using ResNet as a backbone and achieve better results against JPEG compression. Also, the proposed techniques outperform the state-of-the-art methods reported in the literature.