Double JPEG compression with forgery detection

Abstract

Detecting modified images has become increasingly crucial in combating fake news and protecting people's privacy. This is particularly significant for JPEG images, which are widely used online. Tampering with JPEG images often involves recompression using a different quantization table, which alters the histograms of the original image's discrete cosine transform (DCT) coefficients. This study exploits this double compression effect to propose a novel deep learning model that combines a CNN and a stacked residual bidirectional long short-term memory (Bi-LSTM) model that incorporates self-attention mechanisms. A CNN model is initially used to learn the characteristics of DCT coefficients and quantization tables extracted from JPEG files. Subsequently, these features are fed into a stacked residual Bi-LSTM model with an attention mechanism to effectively capture the data's long-term forward and backward relationships. By leveraging the strengths of these diverse techniques, we construct a deep Bi-LSTM with up to five layers, which achieves superior predictive performance compared to existing methods. Our model demonstrates its potential for the robust detection and localization of JPEG forgery.