Training-Free Color-Aware Adversarial Diffusion Sanitization for Diffusion Stegomalware Defense at Security Gateways

Vladimir Frants, Sos S. AgaianarXiv 2025 · DOI · arXiv

Abstract

Generative AI has enabled "coverless" steganography, where secret messages are embedded during the image generation process, bypassing traditional detection. This paper presents Adversarial Diffusion Sanitization (ADS), a training-free defense mechanism for security gateways. ADS leverages a pretrained denoiser to apply targeted perturbations in the diffusion space, disrupting the mathematical pathways used by steganographic decoders. By incorporating quaternion algebra (ADS-QDir), the method ensures high visual fidelity while effectively neutralizing hidden payloads. Experiments show that ADS-QDir achieves a 98.4% decoder failure rate with minimal perceptual distortion.

1. Introduction

Traditional steganography relies on modifying existing cover images, creating statistical anomalies that security tools can detect. However, the advent of generative AI has introduced "coverless" steganography. In this paradigm, an attacker uses a secret message as a seed for a diffusion model to generate a benign-looking image from scratch. Since no original image exists for comparison, traditional detection methods are rendered ineffective. This poses a significant threat in the form of stegomalware, where malicious payloads for command-and-control (C2) operations are hidden within AI-generated content.

2. Adversarial Diffusion Sanitization (ADS)

Rather than attempting the increasingly difficult task of detection, the authors propose an "active warden" approach: sanitizing every incoming image at the security gateway. The proposed method, Adversarial Diffusion Sanitization (ADS), is a training-free approach that disrupts the communication channel between the attacker and the malware receiver.

Figure 1: The ADS framework disrupts the communication channel between an attacker and receiver by sanitizing images at the security gateway.

2.1 The Sanitization Pipeline

ADS operates through a three-step "sandwich" process:

  1. Forward Noising: The input image $x$ is mapped into the diffusion latent space by adding a single step of Gaussian noise.
  2. Adversarial Update: Using a pretrained denoiser (e.g., a standard U-Net) as a proxy for the attacker's decoder, ADS calculates a targeted perturbation. This perturbation maximizes the "reverse-consistency loss," effectively scrambling the mathematical trajectory required for successful message extraction.
  3. One-Step Reverse: The perturbed latent is converted back into the image space, resulting in a sanitized image $x'$.

2.2 Color-Awareness with ADS-QDir

A common issue with per-channel adversarial perturbations is the introduction of "rainbow speckle" artifacts. To maintain visual fidelity, the authors introduce ADS-QDir. This variant utilizes quaternion algebra to treat the Red, Green, and Blue channels as a single coupled entity. By performing the adversarial update in the quaternion domain, the method preserves the color relationships of the original image, ensuring the sanitization is imperceptible to human observers.

3. Experimental Evaluation

The defense was evaluated against Pulsar, a state-of-the-art diffusion steganography scheme, using 64 images generated via the google/ddpm-church-256 backbone. Each image contained a substantial 8192-byte payload.

3.1 Performance Metrics

The effectiveness of the defense was measured by the Decoder Success Rate (DSR), while image utility was assessed using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS).

3.2 Key Results

ADS-QDir demonstrated superior performance compared to traditional sanitization methods like JPEG compression or resizing:

  • High Security: ADS-QDir reduced the decoder success rate to a mere 1.6% (a 98.4% failure rate).
  • Superior Fidelity: This level of security was achieved at a very low distortion level (LPIPS ≈ 0.0022, SSIM = 0.994). In contrast, achieving a similar DSR with JPEG compression required significantly higher distortion, resulting in visible blurring.

4. Conclusion and Future Work

ADS provides a robust, training-free defense against the emerging threat of diffusion-based stegomalware. By shifting the focus from detection to sanitization, security gateways can neutralize hidden payloads without needing prior knowledge of the specific steganographic scheme. While the current implementation introduces a slight computational overhead (83ms per image), future work will focus on optimizing performance for high-traffic environments and extending the framework to other generative modalities such as audio and video.

Summary updated

BibTeX
@misc{frants2025training,
  author = {Vladimir Frants and Sos S. Agaian},
  title = {{Training-Free Color-Aware Adversarial Diffusion Sanitization for Diffusion Stegomalware Defense at Security Gateways}},
  year = {2025},
  eprint = {2512.24499},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.24499},
}
Figure