Robust Manipulation Parameter Estimation Scheme for Post-JPEG Compressed Images using CNN

Abstract

The importance of estimating the manipulation parameters in forged images has increased to grasp the life cycle of the image and gives forensic clues. Although several techniques have been put forth to do this, they are limited to a single alteration in the uncompressed scenario. However, estimation is difficult for multiple manipulations because a new image modification is produced every time. Therefore, it is necessary to repeat the procedure of creating a mathematical framework to determine a parameter estimation. To get around this, we create a deep residual framework that can estimate multiple manipulation parameters in a post-JPEG compressed image with achievable arbitrary precision. Specifically, the framework goes through the steps of noise residual extraction, feature extraction, and classification. To begin, the front-end detector is greatly augmented by the noise residual extraction stage, which adds three residual blocks with skip connections to generate noise residuals by reducing visual features and boosting alteration clues. Further, characteristics of the deep tampering clues are then obtained using a cross-feature learning technique and procured to fully connected layers for classification. To improve forged parameter estimates in the real world, we conduct experiments on three modifications and then apply several quality factors to the results. To top it all off, the proposed MPeR-Net outperforms state-of-theart methodologies in the most crucial instance of lossy post-JPEG compression