[Paper Review] Destruction of Image Steganography using Generative Adversarial Networks

Isaac Corley, Jonathan Lwowski+1 more|arXiv (Cornell University)|Dec 20, 2019

Advanced Steganography and Watermarking Techniques29 references14 citations

TL;DR

This paper proposes the Deep Digital Steganography Purifier (DDSP), a Generative Adversarial Network (GAN)-based model that removes steganographic content from images while preserving high perceptual quality. Evaluated on multiple steganography algorithms and file formats, DDSP achieves the highest steganography destruction rate and best visual fidelity compared to state-of-the-art methods, including successful transfer learning to unseen steganographic techniques and audio steganography.

ABSTRACT

Digital image steganalysis, or the detection of image steganography, has been studied in depth for years and is driven by Advanced Persistent Threat (APT) groups', such as APT37 Reaper, utilization of steganographic techniques to transmit additional malware to perform further post-exploitation activity on a compromised host. However, many steganalysis algorithms are constrained to work with only a subset of all possible images in the wild or are known to produce a high false positive rate. This results in blocking any suspected image being an unreasonable policy. A more feasible policy is to filter suspicious images prior to reception by the host machine. However, how does one optimally filter specifically to obfuscate or remove image steganography while avoiding degradation of visual image quality in the case that detection of the image was a false positive? We propose the Deep Digital Steganography Purifier (DDSP), a Generative Adversarial Network (GAN) which is optimized to destroy steganographic content without compromising the perceptual quality of the original image. As verified by experimental results, our model is capable of providing a high rate of destruction of steganographic image content while maintaining a high visual quality in comparison to other state-of-the-art filtering methods. Additionally, we test the transfer learning capability of generalizing to to obfuscate real malware payloads embedded into different image file formats and types using an unseen steganographic algorithm and prove that our model can in fact be deployed to provide adequate results.

Motivation & Objective

• To address the limitations of traditional steganalysis, which suffer from high false positive rates and poor generalization across image types and steganography algorithms.
• To develop a filtering solution that destroys steganographic content without degrading image quality, avoiding the impracticality of blocking suspicious images.
• To create a model capable of generalizing to unseen steganography algorithms and file formats, including audio, without retraining.
• To optimize steganography destruction while maintaining perceptual image quality through deep learning.
• To provide a practical, deployable solution for real-world network defense against APTs using image-based malware exfiltration.

Proposed method

• The DDSP model is a conditional GAN trained to transform steganographically embedded images into clean, perceptually indistinguishable outputs.
• It uses a generator network to remove hidden data and a discriminator to ensure the output remains visually realistic and free of steganographic artifacts.
• The model is trained on a diverse dataset of images embedded with payloads using various steganography techniques, including LSB, Deep Steganography, and ISGAN.
• Transfer learning is applied by testing the pretrained DDSP model on unseen steganography algorithms (e.g., LSB) and file formats (e.g., PGM, audio) without retraining.
• For audio steganography, input signals are reshaped into 2D matrices, processed by the DDSP model, then restored to 1D vectors and filtered using a Butterworth lowpass and Hanning window to reduce artifacts.
• Performance is evaluated using metrics including Bit Error Rate (BER), Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Universal Image Quality Index (UQI).

Experimental results

Research questions

• RQ1 Can a GAN-based model effectively destroy steganographic content in images while maintaining high perceptual image quality?
• RQ2 How does the DDSP model perform on steganography algorithms and image formats it was not explicitly trained on?
• RQ3 To what extent can the DDSP model generalize to non-image data, such as audio files with steganographic payloads?
• RQ4 Does the DDSP model outperform traditional filtering methods like wavelets and bicubic interpolation in both steganography removal and visual fidelity?
• RQ5 Can the model be deployed without retraining to handle real-world, unseen steganographic payloads and file types?

Key findings

• On the BOSSBase dataset with LSB-embedded images, DDSP achieved a BER of 0.82, MSE of 5.09, PSNR of 41.05, SSIM of 0.98, and UQI of 0.99, outperforming all baselines.
• The pretrained autoencoder and wavelet filter methods showed lower BER (0.78 and 0.53, respectively) but worse image quality, with MSE values of 5.63 and 6935.08, respectively.
• Bicubic interpolation performed poorly, with a PSNR of only 9.83 and MSE of 6763.73, indicating significant image degradation.
• In transfer learning to audio steganography using the VoxCeleb1 dataset, DDSP achieved a BER of 0.67, MSE of 650.12, and PSNR of 37.28, matching the autoencoder and outperforming bicubic interpolation (PSNR: 35.54).
• The wavelet denoising filter slightly outperformed DDSP in PSNR (37.65 vs. 37.28), but DDSP demonstrated superior generalization and adaptability across modalities.
• The results confirm that DDSP generalizes effectively to unseen steganography algorithms and file formats, including audio, without retraining, indicating strong robustness and transferability.