AGASI: A Generative Adversarial Network-Based Approach to Strengthening Adversarial Image Steganography

Area of Science:

• Computer Science
• Information Security
• Artificial Intelligence

Background:

• Steganography is vital for image privacy protection.
• Advanced steganalysis, particularly deep learning models, threatens traditional steganography.
• Existing methods struggle to maintain image quality and robustness against detection.

Purpose of the Study:

• To propose AGASI, a Generative Adversarial Network (GAN)-based approach to enhance adversarial image steganography.
• To improve the robustness of stego-images against sophisticated steganalysis techniques.
• To maintain high-quality secret image extraction while increasing resistance to detection.

Main Methods:

• Utilized an encoder as a generator and a discriminator to form a GAN for adversarial training.
• Implemented a GAN framework to minimize the difference between original and extracted secret images.
• Employed a decoder for effective secret image extraction from stego-images.

Main Results:

• AGASI ensures high-quality secret images and effective extraction.
• The method significantly reduces the accuracy of neural network steganalysis classifiers, inducing misclassifications.
• Achieved an 84.73% misclassification rate under PGD attack, a 23.31% increase over comparable methods.
• Demonstrated increased embedding capacity in the steganography system.

Conclusions:

• AGASI provides a robust solution for image privacy protection against advanced steganalysis.
• The GAN-based approach effectively balances stego-image quality, detection evasion, and embedding capacity.
• AGASI represents a significant advancement in secure steganography techniques.