Forensic Method for JPEG Image Reconstruction

‍ ‍Digital images have been widely used in various online businesses and as judicial evidence. Simultaneously， using popular image editing software， ordinary users can tamper with the image semantics without leaving visual traces. Therefore， identifying the originality and authenticity of digital images has become an urgent application requirement. Image tampering forensics based on metadata has garnered attention due to its high accuracy and minimal computational requirements. However， the emergence of original image reconstruction technologies， exemplified by tools like the MagicEXIF metadata editor， renders the aforementioned methods entirely ineffective. To solve this problem， this study proposes a JPEG original image reconstruction forensics method to detect if the image was reconstructed. By analyzing the original image reconstruction process and the difference in pixel statistical characteristics of the image before and after reconstruction， this study develops a lightweight improvement on the deep learning steganography analysis model steganalysis residual network （SRNet）： it cuts its redundant lower sampling layer to reduce parameters， introduces the channel attention mechanism to improve the ability to extract key features， and uses the knowledge distillation method to further improve the accuracy of the model. Furthermore， by analyzing the influence of reconstruction on different color components， the YCbCr color component is used as the model input to improve the detection performance. To test the performance of the algorithm， we collected image data captured by mobile phones of different brands and models and built a large-scale reconstruction image dataset. The experiment demonstrates that the proposed model outperforms a popular model， even with a significantly reduced number of parameters. For a 512×512 image， the proposed model achieves a detection accuracy exceeding 98% and exhibits strong cross-device generalization capability. Simultaneously， through the application of transfer learning， the proposed method also achieved good generalization for different versions of reconstructed software.