README.md

---
pretty_name: StegBench
task_categories:
  - image-classification
tags:
  - steganalysis
  - steganography
  - digital-forensics
  - deep-learning
  - jpeg
  - png
  - dct
  - lsb
  - cybersecurity
size_categories:
  - 100K<n<1M
---

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

> The image files are intentionally not yet available. This repository currently contains the dataset card, taxonomy, and construction details. The complete StegBench dataset will be publicly released upon publication of the accompanying paper.

## Dataset Summary

StegBench is a benchmark for *multi-class steganalysis*, going beyond a simple cover-vs-stego decision to identify which steganographic algorithm was used to embed a hidden payload. Knowing the embedding algorithm narrows down the relevant analysis and extraction techniques, facilitates forensic investigation, and may help link a suspicious file to a known toolchain.

The dataset is derived from **70,000 cover images** and comprises **525,000 labeled images**, organized into four branches that combine two image formats (JPEG and PNG) with two color spaces (grayscale and color):

- **JPEG Grayscale** — built from BOSSBase 1.01
- **JPEG Color** — built from ALASKA2
- **PNG Grayscale** — built from BOSSBase 1.01
- **PNG Color** — built from Flickr30k (converted to lossless PNG)

Each branch pairs a set of steganographic algorithm classes with a **clean** class and three **noise-perturbation classes** (Gaussian, salt-and-pepper, uniform). The noise classes are not meant to be attributed as if they were steganography — they act as a control group, used to check whether a detector is really picking up on embedding-specific statistical signatures or just reacting to any kind of pixel-level perturbation.

The JPEG branches use a DCT-domain taxonomy (F5, JSteg, Steghide, OutGuess), and the PNG branches use a spatial-domain, LSB-based taxonomy (sequential LSB, pseudo-random LSB via OpenStego, and LSB matching). Baselines trained on StegBench (ConvNeXt-Large, EfficientNetV2-L, SwinV2-Large, and SRNet) reach very high within-domain accuracy, but transferring a model across formats (JPEG ↔ PNG) or across color spaces (grayscale ↔ color) causes a sharp drop in performance — this generalization gap is the main phenomenon the benchmark is designed to expose and measure.

## Dataset Structure

### Branches and classes

| Branch | Source corpus | Resolution | # Classes | Classes |
|---|---|---|---|---|
| JPEG Grayscale | BOSSBase 1.01 | 512×512 | 8 | clean, f5, jsteg, steghide, outguess, gaussian_noise, saltpepper_noise, uniform_noise |
| JPEG Color | ALASKA2 | 512×512 | 8 | clean, f5, jsteg, steghide, outguess, gaussian_noise, saltpepper_noise, uniform_noise |
| PNG Grayscale | BOSSBase 1.01 | 512×512 | 7 | clean, lsb, openstego, lsb_matching, gaussian_noise, saltpepper_noise, uniform_noise |
| PNG Color | Flickr30k | 384×384 | 7 | clean, lsb, openstego, lsb_matching, gaussian_noise, saltpepper_noise, uniform_noise |

Note: the PNG Color branch uses a different resolution (384×384) than the other three branches (512×512), because Flickr30k images have heterogeneous, generally lower native resolution. This is documented as a known confound for cross-format comparisons involving that branch.

### Splits

Every branch is split **70/15/15** (train/validation/test) at the level of *cover images*, using a fixed random seed. All derived variants inherit the split of their source cover image, so no cover image or its derivatives ever appear in more than one split, and every class within a branch has an identical, class-balanced number of train/val/test images. The following counts correspond to the dataset that will be publicly released upon publication of the accompanying paper.

| Branch | Train | Validation | Test | Total cover images |
|---|---|---|---|---|
| JPEG Grayscale | 7,000 | 1,500 | 1,500 | 10,000 |
| PNG Grayscale | 7,000 | 1,500 | 1,500 | 10,000 |
| JPEG Color | 17,500 | 3,750 | 3,750 | 25,000 |
| PNG Color | 17,500 | 3,750 | 3,750 | 25,000 |
| **Total** | **49,000** | **10,500** | **10,500** | **70,000** |

### Planned Directory Structure

```
StegBench/
├── JPEG/
│   ├── grayscale/
│   │   ├── train/{clean,f5,jsteg,steghide,outguess,gaussian_noise,saltpepper_noise,uniform_noise}/
│   │   ├── val/...
│   │   └── test/...
│   └── color/
│       └── ...
└── PNG/
    ├── grayscale/
    │   └── ...
    └── color/
        └── ...
```

## Source Data

- **BOSSBase 1.01** — 10,000 grayscale cover images (512×512), used as the shared source for both grayscale branches.
- **ALASKA2** — native color JPEG photographs from 479 camera models, used for the JPEG Color branch.
- **Flickr30k** — natural color photographs, converted to lossless PNG and resized to 384×384, used for the PNG Color branch.

## Intended Use and Limitations

StegBench is intended as **diagnostic and evaluation infrastructure** for multi-class steganalysis research, not as a plug-and-play production detector. In particular:

- It covers classical, non-adaptive embedding algorithms rather than content-adaptive schemes (e.g. J-UNIWARD, WOW, HUGO).
- It is perfectly class-balanced by construction, which does not reflect real-world deployment priors (where clean images vastly outnumber stego images).
- The algorithm taxonomy is closed-set: models are evaluated only on algorithms seen during training, not on unseen ones.
- Robustness was only tested against JPEG recompression; other post-processing operations (resizing, cropping, format conversion) remain unevaluated.

Full details on dataset construction, embedding parameters, and known limitations are provided in the accompanying paper.

## Citation

If you use StegBench, please cite the accompanying paper. The full citation will be added upon publication.

## License

The license will be announced upon publication of the accompanying paper. Note that the underlying source corpora (BOSSBase, ALASKA2, Flickr30k) carry their own terms, which apply to the corresponding portions of StegBench.