SELF ATTENTION CONVOLUTIONAL NEURAL NETWORK (SACNN) BASED PREDICTOR FOR REVERSIBLE DATA HIDING IN IMAGES

Abstract

The traditional image predictors used in reversible data hiding (RDH) schemes are limited by their inability to capture context from a larger set of image pixels. This work, proposes a new predictor for reversible data hiding, that uses the Self Attention Convolutional Neural Network (SACNN) for improvement in prediction process. With the help of an image division scheme, our method divides a grayscale image into two separate sets. The first set serves as input to the SACNN predictor, which predicts the second set, which will be utilized for data embedding. Our network efficiently captures both local and global dependencies through the use of self attention mechanism in combination with convolutional layers, allowing for accurate pixel prediction and improving overall prediction accuracy. The predictor is trained on over 1000 images randomly taken from the ImageNet dataset. The experimentation and analysis show that the predictor is able to generate a sharper prediction error histogram and can be utilized for achieving a better embedding performance by using it with a suitable scheme in future.