HAG-NET: Hiding Data and Adversarial Attacking with Generative Adversarial Network.

Area of Science:

• Computer Science
• Information Security
• Artificial Intelligence

Background:

• Existing watermarking techniques struggle to protect carrier data from sophisticated steganalysis, particularly those employing adversarial perturbations.
• Deep Neural Networks (DNNs) are highly sensitive to minor perturbations, posing a risk to carrier data integrity in steganographic methods.

Purpose of the Study:

• To propose HAG-NET, a novel watermarking method that jointly trains an encoder, decoder, and attacker to protect carrier data.
• To generate Adversarial Steganographic Examples (ASEs) that are robust against target classification networks, thereby safeguarding the carrier image.
• To ensure the decoder can accurately recover secret data from these protected ASEs.

Main Methods:

• The HAG-NET method employs a jointly trained encoder, decoder, and attacker framework.
• The encoder generates ASEs designed to be adversarial to a target classification network, thus protecting the carrier data.
• The decoder is trained to recover secret information embedded within the generated ASEs.

Main Results:

• HAG-NET achieved an average success rate exceeding 99% in generating ASEs on MNIST and CIFAR-10 datasets.
• ASEs generated by HAG-NET demonstrated enhanced robustness, increasing attack ability by approximately 3.32%.
• Image information entropy measurements confirmed that HAG-NET embeds significantly more information compared to other generative stego examples at similar perturbation levels.

Conclusions:

• HAG-NET offers a robust solution for protecting carrier data in watermarking applications against DNN-based steganalysis.
• The method successfully balances carrier data protection with the embedding and recovery of secret information.
• HAG-NET represents an advancement in secure steganography by leveraging adversarial examples for enhanced data protection.