Research On Image Steganography Based On Semantic Importance

The explosive development of computer network technology and the widespread popularization of mobile smart devices may cause serious information security problems,and steganography is an effective method to ensure information security.Since digital images widely spread in the network,faced with the demand for the secure transmission of massive secret images and increasingly powerful steganalysis techniques,under the premise of ensuring extraction effect at the receiver,steganography algorithms with a larger payload capacity and better steganography effect are required.Different regions of the secret images have different degrees of semantic importance:regions with complex textures and contents have higher semantic importance,and more features need to be preserved for restoration,while regions with simple textures and contents have lower semantic importance,and only a few features are needed to achieve fine extraction effect.Semantic importance can help to remove redundancies that do not affect the extraction effect or that have a limited impact,further enhancing the steganography effect.Therefore,selected from "BUPTTelecom Vision Intelligence Joint Laboratory Project",this thesis proposes an image steganography algorithm based on semantic importance.The main achievements include:(1)Aiming at taking advantage of semantic importance to reduce the difference between the cover images and the embedded ones,for the scene considering the sender and the receiver,Semantic-Weighted Image Steganography Algorithm(SISA)is proposed,which embeds color secret images into cover images of the same size.The network adopts the structure of an autoencoder,and it is composed of the Semantic-based Prep Network,the Hiding Network,and the Reveal Network.The Semanticbased Prep Network analyzes the semantic importance of each pixel in the secret images and obtains the semantic importance map.The secret image features are then weighted and cropped according to the semantic importance map.Experimental results show that the average pixel deviation between the embedded images and the cover images is reduced by up to about 36.73%using the proposed algorithm,compared with the existing algorithms that do not apply semantic weighting.Therefore,transparency is improved.(2)Aiming at taking advantage of semantic importance to reduce the possibility of secret images being detected by the third-party,based on SISA,for the scene considering the sender,the receiver,and the third-party,Semantic-Weighted and Secure-Enhanced Image Steganography Algorithm(SSISA)is proposed,which embeds color secret images into cover images of the same size.The network adopts the structure of a generative adversarial network,which is composed of the Generative Network and the Discrimination Network.The Discriminant Network is introduced to simulate the steganalysis process of the third party for adversarial training.Experimental results show that the steganalysis detection rate is reduced by up to about 6.75%using the proposed algorithm,compared with the existing algorithms that do not apply security enhancement.Therefore,security is further improved.