Adversarial Data Hiding in Digital Images.

Area of Science:

• Computer Vision
• Machine Learning
• Cybersecurity

Background:

• Generative Adversarial Networks (GANs) are explored for image privacy and authenticity.
• Existing methods often compromise image quality or data concealment for adversarial effects.
• Visible watermarks limit the effectiveness of adversarial perturbation and data hiding.

Purpose of the Study:

• To propose a true data hiding method with adversarial effects for high-quality image generation.
• To enhance image privacy and authenticity simultaneously.
• To develop a GAN-based approach for generating covert adversarial samples.

Main Methods:

• Utilizing GANs to precisely select data hiding areas by limiting modification strength, preserving image fidelity.
• Employing a genetic algorithm to explore decision boundaries for improved attack effects.
• Detecting "sensitive pixels" to place discontinuous perturbations for aggressive covert adversarial samples.

Main Results:

• The proposed method generates stego-images with good visual quality.
• The generated adversarial samples demonstrate a significant attack effect.
• Achieved simultaneous data hiding and adversarial perturbation without visible watermarks.

Conclusions:

• This work presents the first known method using covert data hiding to generate adversarial samples based on GANs.
• The approach successfully balances image quality, data concealment, and adversarial robustness.
• Offers a promising direction for secure image data hiding and adversarial attack research.