Research shows deep image hiding models are vulnerable to cross-model secret extraction

1. 23:05
   arXiv: Alignment, Deception & Oversight

   Your secrets are being exposed: deep image hiding models are surprisingly easy to attack

   The authors demonstrate a cross-model extraction attack against deep image hiding (steganography) models, in which an attacker can extract and identify the secret content without knowing which model hid the image. The authors built an evaluation dataset from steganographic images generated by 11 hiding models, covering the invertible neural network and autoencoder families, CNN, Transformer, and Mamba backbones, and spatial-domain and frequency-domain methods; a single HiNet decoder trained in the conventional way can extract secret content from these models unseen in these settings, and both humans and machines can identify its content. Frequency-domain analysis shows that most models prefer mid-to-high-frequency embedding, and the authors accordingly strengthen cover texture and suppress secret texture during training to enhance cross-model extraction. The attack is also effective against out-of-distribution secrets and multi-image hiding, and can expose private content hidden by the steganography-based privacy image restoration method AIRPNet; the authors say the decoder for each tested model has this general extraction capability to varying degrees, and code and dataset will be open-sourced.

Current attention 33·Peak in the comparable period 35 (Nov 13 17:00)·Change over the comparable past 24 hours -2%