RoSteALS: Robust Steganography using Autoencoder Latent Space

Tu Bui, Shruti Agarwal, Ning Yu, John P. CollomosseWorkshops @ CVPR 2023 · DOI · OpenReview

TL;DR

This paper introduces RoSteALS, a novel steganography technique that embeds secret data directly into the latent space of pre-trained autoencoders. This approach achieves robust secret recovery and high image quality with a surprisingly small, easy-to-train secret encoder, and even enables new applications like cover-less and text-based steganography.

What Problem Does This Paper Address?

In an era of deepfakes and rampant misinformation, verifying the origin and integrity of digital images is crucial. While methods like digital watermarking can embed identifiers, existing techniques often struggle with a trade-off: either they preserve image quality at the expense of robustness against modifications, or they are robust but introduce noticeable artifacts and are complex to train. Furthermore, many methods require learning the entire image distribution, which is computationally expensive. The authors aim to bridge this gap by proposing a steganography method that is robust, preserves image quality, is computationally efficient, and can be adapted to emerging generative AI applications.

Methods

RoSteALS leverages the "free knowledge" from pre-trained, frozen autoencoders (specifically VQGAN in their experiments). The core innovation is injecting the secret data directly into the latent space of the autoencoder.

  • Latent Steganographic Embedding: Instead of learning the complex image distribution, RoSteALS uses a small, lightweight "secret encoder" (around 300k parameters) to map a secret bit string to a small offset ($\delta$). This offset is then added to the latent representation ($z$) of a cover image before it's decoded by the autoencoder's generator. This decouples the secret embedding from learning the cover image distribution.
  • Robust Secret Recovery: To ensure robustness against image perturbations (like noise, compression, etc.), the authors introduce a noise model between the autoencoder's generator and the secret decoder. This noise model simulates various real-world degradations, forcing the secret decoder to learn to recover the secret even from corrupted stego images.
  • Cover-less and Text-Based Steganography: The latent space manipulation allows for novel applications. By learning a mapping from a simple distribution (like Gaussian noise) to the latent space, they can generate stego images without needing an explicit cover image. They also extend this to text-based steganography by conditioning a latent diffusion model (LDM) on text prompts to generate latent codes that embed secrets.

The training objective balances image quality (using pixel and perceptual losses like LPIPS) and secret recovery (using binary cross-entropy loss). A crucial aspect of their training strategy is a curriculum learning approach that prioritizes secret recovery early on, then gradually introduces perturbations and ramps up the quality loss weight.

Datasets

  • Training: MIRFlickR dataset (100K images for training, 1K for validation).
  • Evaluation:
    • CLIC (Content-based Image Retrieval and Compression) dataset.
  • Met-Face dataset.
  • Stock1K dataset.
  • Perturbations: ImageNet-C dataset (14 types of common image corruptions).

Key Insights and Ideas

  • Latent Space as a Steganographic Canvas: The paper convincingly demonstrates that the latent space of powerful pre-trained autoencoders is a surprisingly effective and robust medium for hiding information, without needing to understand the intricacies of the image pixels themselves.
  • Decoupling for Efficiency and Robustness: By freezing the autoencoder and only training a small secret encoder/decoder, RoSteALS achieves remarkable efficiency and robustness. This separation allows the secret embedding to focus on its task without being hindered by the complexity of learning image distributions.
  • Adaptability to Generative AI: The latent space approach naturally extends to cover-less and text-based steganography, showcasing the versatility of the method in the context of modern generative models like diffusion models.
  • Curriculum Learning for Robustness: The staged training approach, prioritizing secret recovery before introducing noise, is key to achieving high robustness without sacrificing image quality.

Key Results

  • Image Quality: RoSteALS achieves comparable image quality to other state-of-the-art methods across three benchmarks (CLIC, Met-Face, Stock1K), often outperforming them in specific metrics like SIFID on Met-Face.
  • Secret Recovery: RoSteALS demonstrates near-perfect secret recovery on clean data and significantly outperforms baselines (including StegaStamp and dwtDctSvd) under various ImageNet-C perturbations, especially on challenging corruptions like blurring and heavy enhancement effects.
  • Efficiency: The secret encoder is very small (300k parameters), making it easy and fast to train. Training time is also significantly less than some other methods like SSL or RivaGAN.
  • Cover-less/Text-based: RoSteALS achieves high bit accuracy (>99% for cover-less, >98% for text-based) when integrated with LDM models, demonstrating the viability of these novel applications.

Limitations and Future Work

  • Autoencoder Dependency: RoSteALS inherits limitations from its underlying autoencoder. If the autoencoder struggles to reconstruct fine details or small text, RoSteALS will also be affected, even if the secret embedding itself is robust.
  • Quality-Recovery Trade-off: While RoSteALS offers a good balance, increasing secret length or robustness requirements can still lead to a slight decrease in image quality, though the trade-off is manageable.
  • Future Work: The authors suggest exploring more powerful autoencoders, novel fine-tuning strategies for autoencoders to benefit steganography, and further investigating the latent space for more advanced steganographic applications.

Summary updated

Abstract

Data hiding such as steganography and invisible watermarking has important applications in copyright protection, privacy-preserved communication and content provenance. Existing works often fall short in either preserving image quality, or robustness against perturbations or are too complex to train. We propose RoSteALS, a practical steganography technique leveraging frozen pretrained autoencoders to free the payload embedding from learning the distribution of cover images. RoSteALS has a lightweight secret encoder of just 300k parameters, is easy to train, has perfect secret recovery performance and comparable image quality on three benchmarks. Additionally, RoSteALS can be adapted for novel cover-less steganography applications in which the cover image can be sampled from noise or conditioned on text prompts via a denoising diffusion process. Our model and code are available at https://github.com/TuBui/RoSteALS.
BibTeX
@misc{bui2023rosteals,
  author = {Tu Bui and Shruti Agarwal and Ning Yu and John P. Collomosse},
  title = {{RoSteALS: Robust Steganography using Autoencoder Latent Space}},
  year = {2023},
  eprint = {2304.03400},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.03400},
}
Figure