Research On Image Deep Steganography Method Based On Security Measuremen

Nowadays,the development of information intelligence greatly contributes to the growth of the amount of information in networks,and the security of information transmission has been widely emphasized.Meanwhile,how to ensure the security of information transmission in untrustworthy communication channels has become an important demand.Steganography is an important application in the field of covert communication,which uses multimedia carriers to hide secret information.Steganography aims to achieve highly secure transmission without the suspicion of malicious monitoring parties,and has important theoretical significance and applicability for secure communication.With the rapid development of deep learning-based steganalysis algorithms,steganographic security faces many challenges.In deep learning scenarios,the automatic construction of a steganography algorithm can be regarded as the joint learning process of both the steganography module and the security metric module.In particular,the steganography module learns how to embed information to maximize steganographic security on the guidance of the security metric module.However,there are the following issues with the existing methods.1)The security metric modules are limited by the empirical structure and lack of robustness in the design of the network structure and the optimization approach,respectively,making the steganography module vulnerable to being trapped in a local optimum when learning the steganography cost models.2)The steganography modules are limited in steganography mode by the hiding capacity of the cover images.Besides,as for the overall optimization approach,the steganography module is restricted to a specific security metric module,resulting in the constructed steganography algorithm lacking generalization ability in security.3)In terms of overall optimization objectives,the steganography modules are limited to specific security metric modules,resulting in the lack of generalization of the constructed steganography algorithms in security.Focusing on the above issues,the following research contents are conducted to achieve a general researching goal: enhancing the security of the deep image steganography algorithms,while making the high security generalizable.First,at the level of the security metric module,an adaptive distortion-aware structure and an adversarial robustness-based optimization strategy are designed.The proposed approaches can enhance the performance of the security metric module,so that enhance the security of the constructed steganographic cost model in the steganography module during learning.Next,at the level of the steganography module,a novel cover reconstructive steganography approach is designed.Meanwhile,a multimodal security metric module is formulated,according to the characteristics of the cover reconstructive steganography approach,for optimizing the security of the cover images.Finally,at the level of the optimization manner between the security metric module and the steganographic module,a meta-learning-based optimization strategy is proposed to guide the construction of a steganography module with better generalizability in security,by learning a more generalized security metric module.The main research contents,research results,and innovations are as follows.(1)An adaptive security metric-based cost-minimizing steganography approach is proposed.In existing steganographic cost learning methods,the security metric module usually extracts steganographic features using empirical high-pass filter kernels.Those kernels are used for establishing the awareness of subtle modifications produced in the steganographic module by the steganographic cost model.However,in the training scenarios with different embedding capacities,it will degrade the performance of the security metric module,since the lack of adaptive awareness of specific high-pass filtering kernels for different levels of subtle modifications.Thus,it limits the steganographic cost model to build higher security.To solve the above issues,an adaptive aware convolutional layer is designed for the security metric module based on the constrained convolutional structure and the channel attention mechanism.The proposed structure can automatically learn high-pass filter kernels by weakly-constrained rules to achieve adaptive awareness of different levels of subtle modifications,thus enabling the security metric module to extract more discriminative steganographic features.Experimental results demonstrate that a more comprehensive security metric module can be built under the guidance of this module,to enhance the steganographic security.(2)A robust security metric-based cost-minimizing steganography approach is proposed.Compared with the standard vision tasks,the sample features of the stego image and the cover image are too similar in the steganographic task.This is not conducive to the security metric module to establish high robustness,which will mislead the steganography module to construct a sub-optimal steganographic cost model.To solve the above issues,an optimization approach based on adversarial training is proposed to enhance the robustness of the security metric module.Furthermore,a diversified inverse adversarial training strategy and a diversified steganographic features extraction structure are proposed to provide more diversified adversarial sample features for the security metric module.By efficiently expanding the sample space of stego images,the robustness of the security metric module is improved,which facilitates the steganography module to build the steganographic cost models with higher security.Extensive experimental results show that a more secure steganographic cost model can be built using fewer training samples since the diversity of stego images is effectively extended.(3)A cover reconstructive steganography approach based on multimodal security metrics is proposed.In existing steganographic algorithms,the steganography module finds embedding regions with complex texture on the cover image,to hide the secret information.However,this passive form of modification restricts the cover image to its original hiding capability.To improve the security of the cover image,a novel cover reconstructive steganography approach based on the multimodal conditional adversarial network with the motivation of proactively generating embedding regions.And the multimodal security metric module is also designed according to the characteristics of cover reconstruction.Under the guidance of the multimodal security metric module,the steganography module reconstructs objects with rich textures that match the original image distribution and semantics on the cover image according to the information to be embedded.Then,the generated object is used as the embedding candidate region for enhancing the security of the cover image,to enhance the security of the generated stego images.(4)A meta-security metric-based fully automated steganography approach is proposed.Some existing fully automated steganographic approaches employ steganalysis-based metric networks and adversarial loss functions to construct security metric modules,and to guide the learning of steganographic modules.However,these approaches focus only on maximizing security performance on a specific security metric module,resulting in poor generalization of the security on the steganographic module.To address this security issue,learning a metasecurity metric module to guide the learning of a flow-based hiding network is proposed,in which the meta-security metric module consists of a frequency channel attention-based metric network and a contrastive loss function.The overall learning processes form a metaoptimization rule in a bi-level manner.In the inner optimization,the steganography module learns to maximize security performance on the meta-security metric module,while in the outer optimization,the meta-security metric module maximizes its adaptability of security measurement based on multiple known security metric modules.The extensive experimental results demonstrate the proposed approach can resist unknown steganalysis tools,and achieve promising generalization in security.