نوع مقاله : مقاله پژوهشی
نویسندگان
1 استادیار، گروه مخابرات، دانشگاه جامع امام حسین(ع)، تهران، ایران
2 دانشیار، گروه مخابرات، دانشگاه تهران، تهران، ایران.
3 دانشیار، گروه رمز و امنیت، دانشگاه جامع امام حسین(ع)، تهران، ایران
چکیده .
کلیدواژهها
عنوان مقاله [English]
نویسندگان [English]
Steganography of a text across multiple images enhances security and prevents attackers from accessing the hidden message; moreover, generating multiple images with artificial intelligence also facilitates the availability of images for embedding each segment of the hidden text. This paper introduces a steganography system based on Generative Adversarial Networks that uses a three-component Encoder-Decoder-Discriminator architecture, the encoder extracts messages from steganographic images, and the CIFAR-10 dataset is employed for evaluation. This general method of “Generator and Encoder” combines a cross-entropy adversarial loss to fool the discriminator and a message loss to recover the message with the aid of the encoder. The main objective is to reduce the visible alterations in steganographic images compared to the original (Real) images and to enhance the recoverability of the message by the encoder, along with qualitative and quantitative evaluations such as SNR and BCE/MSE for message recovery. Experimental results show that, with proper hyperparameter tuning, messages can be embedded in images with minimal noise, and the message recovery exhibits low error and a desirable SNR. In the proposed method, at epoch 50, the SNR is approximately 34.9 dB. In this case, Loss_D is approximately 0.7 and Loss_G is approximately 1.2, which are suitable values for cross-entropy. The value of LossMsgLossMsg is approximately 0.03, which is considered excellent. It should be noted that the implementation in this article was performed using Python.
کلیدواژهها [English]