giaupham/fae-hybrid-unet-steganography

Frequency-Aware Embedding Hybrid U-Net

This repository releases the PyTorch robust-best checkpoints for the manuscript Frequency-Aware Embedding Hybrid U-Net: Robust Image Steganography Against Geometric and Compression Attacks.

Released Checkpoints

Both checkpoints use seed 456, 256 x 256 input images, a 1,024-bit barcode payload, and 8 x 8 barcode modulation with 64-fold spatial redundancy per bit. This equals 0.015625 bits per spatial pixel, or 0.005208 bits per RGB channel when the denominator includes all three image channels.

Paper Results

The BOSSBase and MS-COCO rotation results differ substantially. Claims about geometric robustness should remain dataset qualified.

Intended Use

These checkpoints support research reproducibility and non-production evaluation of robust image steganography. The released demo is intended to make the embedding, attack, residual, and decoding workflow inspectable.

Limitations

• —Payload capacity is 1,024 bits at 256 x 256 resolution.
• —The barcode payload differs from pixel-level payload baselines that embed 65,536 bits.
• —The spatial transformer estimates affine transformations and does not guarantee recovery under perspective or non-rigid deformation.
• —JPEG Q30 and combined compression-plus-geometry attacks remain difficult.
• —The steganalysis evaluation in the manuscript is preliminary and should be extended with full-scale detectors.
• —MS-COCO rotation robustness is weaker than BOSSBase rotation robustness.

Citation

@article{pham2026frequencyaware,
  title   = {Frequency-Aware Embedding Hybrid U-Net: Robust Image Steganography Against Geometric and Compression Attacks},
  author  = {Pham, Ngoc-Giau and Tran, Hong-Ngoc and Vo, Phuoc-Hung},
  year    = {2026},
  note    = {Manuscript submitted for peer review}
}

License

The checkpoint files are provided for academic research and evaluation. See LICENSE before redistribution or commercial use.