A Novel Steganography Method for Character-Level Text Image Based on Adversarial Attacks

The Internet has become the main channel of information communication, which contains a large amount of secret information. Although network communication provides a convenient channel for human communication, there is also a risk of information leakage. Traditional image steganography algorithms use manually crafted steganographic algorithms or custom models for steganography, while our approach uses ordinary OCR models for information embedding and extraction. Even if our OCR models for steganography are intercepted, it is difficult to find their relevance to steganography. We propose a novel steganography method for character-level text images based on adversarial attacks. We exploit the complexity and uniqueness of neural network boundaries and use neural networks as a tool for information embedding and extraction. We use an adversarial attack to embed the steganographic information into the character region of the image. To avoid detection by other OCR models, we optimize the generation of the adversarial samples and use a verification model to filter the generated steganographic images, which, in turn, ensures that the embedded information can only be recognized by our local model. The decoupling experiments show that the strategies we adopt to weaken the transferability can reduce the possibility of other OCR models recognizing the embedded information while ensuring the success rate of information embedding. Meanwhile, the perturbations we add to embed the information are acceptable. Finally, we explored the impact of different parameters on the algorithm with the potential of our steganography algorithm through parameter selection experiments. We also verify the effectiveness of our validation model to select the best steganographic images. The experiments show that our algorithm can achieve a 100% information embedding rate and more than 95% steganography success rate under the set condition of 3 samples per group. In addition, our embedded information can be hardly detected by other OCR models.