The Effect of the JPEG Implementation on the Cover-Source Mismatch Error in Image Steganalysis

The Effect of the JPEG Implementation on the Cover-Source Mismatch Error in Image Steganalysis

연구 분야: Strategies

학회: 2022 30th European Signal Processing Conference (EUSIPCO)

One of the challenges in steganalysis based on machine learning is cover-source mismatch (CSM). It occurs when a detector is trained on cover and stego pairs generated from one data source and then applied to data of unknown type from a different source. The mismatch leads to an increase of the classification error, which corresponds to a loss of detection performance. This paper studies how much the implementation of the JPEG compression and decompression contributes to the CSM of color image steganalysis. Specifically, it studies the differences between libjpeg versions 6b and 7. The impact on the CSM is measured for J-UNIWARD embedding using JPEG and spatial rich models with an ensemble classifier. It is observed that the use of different libjpeg versions may cause CSM. The error is large when different libjpeg versions are used for feature extraction, and barely measurable when different versions are used for the generation of cover and stego pairs.