Spotlight on Deception: Uncovering Hidden Manipulation

Today's digital technology has made the creation of fraudulent photographs surprisingly easy, given the wide availability of advanced photo editing software. This accessibility has led to a widespread issue of image forgery, where pictures can be altered or completely fabricated with ease, even by individuals with limited technical knowledge. As a result, distinguishing between genuine and manipulated images has become increasingly challenging. Addressing the widespread problem of real-life image manipulation, which frequently includes various alterations, we introduce ManTra-Net, a cohesive deep neural network design. ManTra-Net stands out from other solutions by providing a comprehensive approach that combines detection and localization seamlessly, eliminating the need for extra pre and post-processing. Operating as a fully convolutional network, ManTra-Net is adept at processing images of different dimensions and an array of forgery types, encompassing image splicing, copy-move attacks, image retouching, and other unidentified manipulations. The proposed solution utilizes a self-supervised learning task aimed at recognizing traces of image manipulation by classifying more than 300 different manipulation types. Moreover, identifying forged areas within an image is approached as a task of detecting irregularities within a specific region. To tackle this, a Z-score feature is employed to detect local irregularities, and a method leveraging Long Short-Term Memory (LSTM) is implemented to assess these irregularities locally.