Deep Learning Based Multi-level Image Security Estimation And Its Applications

With the development of Internet and multimedia technologies,image has become an important carrier to transmit information in social media.Meanwhile,image-based information hiding technologies have also attracted more and more attention.Security,which refers to the inability by unauthorized users to have access to the hidden information,is fundamental to the information hiding technologies.However,due to the complexity of image contents,the problem on how to effectively evaluate and estimate image security remains to be solved.Recently,deep learning based methods have been widely applied to different multimedia problems,such as salient object detection,object retrieval and semantic segmentation.These methods have made significant progress compared to the traditional methods,which extract the manually defined features.Hence,the application of deep learning strategy provides a new solution to the image security estimation problem.In this paper,we propose a deep learning based multi-level image security estimation method,which combines both local security and instance security.According to the similarity of appearance between local security maps and saliency maps,especially in complex image contents,such as textures and contours,we first formulate the local security estimation problem as a saliency detection task,and estimate the local security maps from the corresponding images with a learning based method.Then,we quantitatively evaluate the instance security of each image according to the estimated local security map.The proposed image security estimation method can be applied to different information hiding technologies,including image steganography and steganographer detection.We further propose a steganographic simulation method combining the proposed multi-level image security estimation,and a steganographer detection method combining the proposed multi-scale local security estimation.Specifically,the proposed steganographic simulation method first ranks the cover images according to the instance security values,and select images with higher instance security values to further embed secret information.Then,it calculates the modification maps according to the corresponding local security maps,and generates the stego images.Besides,the proposed steganographer detection method first integrates the multi-scale estimated local security maps into the feature maps to extract steganalytic features.Then,the steganographer is detected via a novel Gaussian vote method with the extracted steganalytic features.The proposed deep learning based multi-level image security estimation method and its applications on steganography and steganographer detection have been validated on the public dataset.The experimatal results prove the effectiveness of the proposed image security estimation method and the superiority of its applications.