This paper explores image modeling from the frequency space and introduces DCTdiff, an end-to-end diffusion generative paradigm that efficiently models images in the discrete cosine transform (DCT) space. We investigate the design space of DCTdiff and reveal the key design factors. Experiments on different frameworks (UViT, DiT), generation tasks, and various diffusion samplers demonstrate that DCTdiff outperforms pixel-based diffusion models regarding generative quality and training efficiency. Remarkably, DCTdiff can seamlessly scale up to 512$\times$512 resolution without using the latent diffusion paradigm and beats latent diffusion (using SD-VAE) with only 1/4 training cost. Finally, we illustrate several intriguing properties of DCT image modeling. For example, we provide a theoretical proof of why `image diffusion can be seen as spectral autoregression', bridging the gap between diffusion and autoregressive models. The effectiveness of DCTdiff and the introduced properties suggest a promising direction for image modeling in the frequency space. The code is at https://github.com/forever208/DCTdiff.

Analytical estimators of the steganographic change rate in images, such as WS steganalysis, often operate on the noise residual. The residual can be obtained by estimating the cover content with pixel predictors and subtracting it from the image under analysis. In recent years, we have witnessed the success of new deep learning-based denoisers, such as U-Net, in various fields of image processing. In this study, we revisit WS steganalysis using a U-Net variant as a drop-in replacement for the linear filters originally proposed for cover prediction. A novel property of this U-Net variant is its hand-crafted loss function, which ensures that when predicting from stego images, the prediction errors are uncorrelated with the stego noise, an assumption required by WS steganalysis. Improving especially in the textured regions, the proposed predictor produces accurate and consistent change rate estimates. When used as a detector, our model significantly reduces false positives and thus potentially sets a new baseline for LSB replacement steganalysis.

Operational steganalysis contends with a major problem referred to as the cover-source mismatch (CSM), which is essentially a difference in distribution caused by different parameters and settings over training and test data. Despite it being of fundamental importance in an operational context, the CSM problem is often overlooked in the literature. With the goal to increase the visibility of this problem and attract the interest of the community, the present paper proposes a systematic review of the literature. It summarizes gathered knowledge and major open questions over the last 20 years of active research on CSM: terminology, methods of measurement, known causes, and mitigation strategies. Over 100 papers exploring, mitigating, assessing, or discussing steganalysis under train-test mismatch were collected by sampling scholar databases, and tracing references, cited and generated. For image steganalysis, the literature provided enough evidence to quantify the impact of causes, and the effectiveness of mitigation strategies.

Introduced in 1991, libjpeg has become a well-established library for processing JPEG images. Many libraries in high-level languages use libjpeg under the hood. So far, little attention has been paid to the fact that different versions of the library produce different outputs for the same input. This may have implications on security-related applications, such as image forensics or steganalysis, where evidence is generated by tracking small, imperceptible changes in JPEG-compressed signals. This paper systematically analyses all libjpeg versions since 1998, including the forked libjpeg-turbo (in its latest version). It compares the outputs of compression and decompression operations for a range of parameter settings. We identify up to three distinct behaviors for compression and up to six for decompression.