DocSafe: Toward Practical Print-Proof Image Steganography via Frequency Decomposition and Covariance Alignment

Data hiding techniques such as robust steganography and invisible watermarking are important for applications including aesthetic stamps, copyright protection, privacy-preserving communication, and content provenance. However, existing print-proof steganography methods often struggle to embed messages in small image regions or decode them from very low-resolution printed outputs. In this paper, we introduce DocSafe, a framework that embeds binary messages (up to 256 bits) into images while producing stamp-like outputs robust to digital and physical distortions. The method enables region-aware embedding using spatial masks, analyzes the impact of wavelet sub-band frequencies for improved robustness, and investigates covariance-based losses, identifying Stein divergence for the encoder and AIRM loss for the decoder as the most effective configuration. A lightweight architecture is further proposed to enhance both imperceptibility and decoding accuracy. We achieve 100% bit accuracy from printed images as small as 2×2 cm in the face dataset and at 3×3 cm on the object dataset, demonstrating state-of-the-art performance in print-proof robust steganography, even when messages are embedded in specific image regions using spatial masks. DocSafe also improves visual quality (ColorHisto 0.05, FID 4.11) and maintains strong decoding accuracy under various digital noise conditions.

Document can be accessed here.