BASN -- Learning Steganography with Binary Attention Mechanism

Yang Yang

Abstract

Secret information sharing through image carriers has gained significant research attention due to the ubiquity of digital imagery. However, the rise of convolutional neural networks (CNNs) presents a new challenge: automated steganalysis and task-specific feature distortion. This paper introduces the Binary Attention Steganography Network (BASN), which utilizes a dual-attention mechanism to maintain high security and payload capacity while minimizing the impact on downstream AI tasks.

Introduction

Traditional image steganography primarily aims to hide data from the Human Visual System (HVS). However, in the modern digital landscape, images are increasingly processed by automated systems. Standard steganographic techniques, while visually imperceptible, often introduce subtle pixel variations that significantly distort the feature maps used by CNNs for tasks like image classification.

The Challenge of AI-Driven Steganalysis

A critical vulnerability in existing methods is "feature distortion." Even if an image appears unchanged to a human, minor modifications can lead to catastrophic failures in machine vision. For instance, applying Least Significant Bit (LSB) Matching can cause a robust classifier like ResNet to misidentify a goldfish as a proboscis monkey. This sensitivity makes traditional steganography easily detectable by neural-network-automated steganalysis tools.

Traditional LSB-M steganography can cause significant feature distortion, leading a classifier to misidentify a goldfish as a proboscis monkey despite no visible changes.

Methodology: The Binary Attention Mechanism

BASN addresses these challenges through a Binary Attention Mechanism that optimizes data embedding for both human and machine observers.

1. Image Texture Complexity (ITC) Model

To satisfy the HVS, the ITC model identifies texture-rich regions (e.g., grass, fur) where pixel alterations are naturally masked. The authors utilize Variance Pooling to analyze local patches, preventing global shifts in brightness or contrast that often plague standard variance-based methods.

2. Minimizing Feature Distortion (MFD) Model

The MFD model ensures compatibility with AI tasks. It analyzes the feature maps of a specific task network (such as ResNet-18) to identify pixels with the least influence on high-level classification features. By targeting these "low-impact" pixels, BASN minimizes the risk of triggering automated detection.

Fusion and Inference

The outputs of these models are combined into a final "Embedding Attention" map using a fusion strategy (mean or minimum). During inference, BASN employs Least Significant Masking (LSM) and Permutative Straddling to further obfuscate the hidden data and resist statistical steganalysis.

Experimental Results

The model was evaluated using the ImageNet (ILSVRC2012) dataset, with a compressed version of The Complete Works of William Shakespeare serving as the payload.

  • Capacity: BASN achieved high payload capacities ranging from 2.0 to 3.89 bits per pixel (bpp), significantly outperforming traditional secure thresholds (typically < 0.4 bpp).
  • Task Integrity: Despite high payloads, feature map distortion remained as low as 2%, ensuring that downstream AI tasks remained accurate.
  • Security: Testing against state-of-the-art steganalysis tools (e.g., StegExpose, Yedroudj-Net) showed detection rates near random guessing, indicating high resistance to discovery.

Conclusion

BASN represents a shift in steganographic design, treating the problem as a three-way competition between the sender, the human observer, and the machine classifier. By leveraging binary attention and variance pooling, the framework achieves a balance of high capacity and robust security in an AI-integrated world. Future work will explore extending this task-specific attention to object detection and facial recognition systems.

Summary updated

Abstract

Secret information sharing through image carrier has aroused much research attention in recent years with images' growing domination on the Internet and mobile applications. However, with the booming trend of convolutional neural networks, image steganography is facing a more significant challenge from neural-network-automated tasks. To improve the security of image steganography and minimize task result distortion, models must maintain the feature maps generated by task-specific networks being irrelative to any hidden information embedded in the carrier. This paper introduces a binary attention mechanism into image steganography to help alleviate the security issue, and in the meanwhile, increase embedding payload capacity. The experimental results show that our method has the advantage of high payload capacity with little feature map distortion and still resist detection by state-of-the-art image steganalysis algorithms.
BibTeX
@misc{yang2019basn,
  author = {Yang Yang},
  title = {{BASN -- Learning Steganography with Binary Attention Mechanism}},
  year = {2019},
  eprint = {1907.04362},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.04362},
}
Figure