CNN Based Image Resizing Detection and Resize Factor Classification for Forensic Applications

This paper investigates the forensic problem of resizing detection along with determining the factor by which the image underwent resizing in the uncompressed scenario. In many forensic applications, there is a need to understand the life history of the image under analysis since multimedia data is slowly becoming admissible as evidence in the court of law. The work reported in the paper is a novel attempt in this direction where a convolutional neural network is utilized with twin objectives. Firstly, to detect the presence of resizing by capturing the forensic clues left behind by the resizing operation and secondly, to determine by what factor the uncompressed image was resized which is a blind estimation. Experimental simulation utilizing the raise dataset shows very high accuracy scores for the proposed method. To check the robustness of the proposed network, an adversarial attack, namely Carlini and the Wagner attack, is a white-box attack aiming towards system failure. To the best of our knowledge, such a detailed forensic analysis with one of the adversarial attacks to validate the security of the proposed method has never been reported in the literature and is the novel contribution of the work.