StegBench: A Dual-Branch Benchmark Dataset for Multi-Class Steganalysis in JPEG and PNG Formats Using Deep Learning

The increasing complexity of steganographic techniques used to conceal information in digital content constitutes a major challenge for cybersecurity and digital forensics. Although deep learning has achieved strong performance in binary steganalysis, generalization across embedding algorithms, image formats, and datasets remains limited, highlighting the need for controlled benchmarks that can systematically evaluate such distribution shifts. We propose StegBench, a reference dataset for multi-class steganalysis for JPEG and PNG formats, with public release on Hugging Face planned upon publication. Beyond simply flagging a suspicious file for quarantine, identifying the specific embedding algorithm provides additional information that can support further analysis. In particular, it reduces the number of extraction methods that need to be considered when attempting to recover the hidden payload. Moreover, once the payload is recovered, this information can help associate the analyzed file with known steganographic tools or toolchains, supporting threat intelligence and the correlation of related security incidents. StegBench is constructed from 70,000 source cover images and comprises 525,000 labeled images distributed across four complementary branches, structured on grayscale and color segments for each format: the JPEG branches are derived from the BOSSBase 1.01 and ALASKA2 sets, while the PNG branches are built on the basis of BOSSBase 1.01 and Flickr30k. The JPEG branch uses an eight-class taxonomy comprising unaltered images, four DCT-domain steganographic algorithms (F5, JSteg, Steghide, and OutGuess), and three types of noise-based perturbations. The noise-based categories serve as adversarial foils, allowing us to assess whether the detectors learn genuine embedding signatures rather than merely responding to generic perturbation energy. Accordingly, these categories are included for robustness evaluation and are not intended for steganographic algorithm attribution. The PNG branch integrates LSB steganographic methods, along with the same noise categories, thus enabling direct comparison between the mentioned formats. The evaluation includes EfficientNetV2-L, ConvNeXt-Large, and SwinV2-Large, together with SRNet as a specialized steganalysis baseline; the general-purpose backbones were additionally evaluated with a fixed SRM high-pass front-end to preserve steganalysis-relevant residual information. The ConvNeXt-Large model recorded the best results, reaching an accuracy of 99.60% on the color JPEG dataset. Cross-domain evaluations confirm the expected collapse when transferring between the DCT-domain (JPEG) and spatial-domain (PNG) embedding regimes, and additionally reveal a comparably severe degradation when transferring across color distributions (grayscale vs. color) within the same format, which is a less predictable result indicating that models rely on source- and channel-specific statistics rather than transferable steganographic signatures. These results demonstrate the value of dataset diversity for systematically exposing generalization limitations and provide a controlled benchmark for evaluating future domain-robust steganalysis methods.