Reinforcement learning of non-additive joint steganographic embedding costs with attention mechanism

作     者：Weixuan TANG Bin LI Weixiang LI Yuangen WANG Jiwu HUANG 

作者机构：Institute of Artificial Intelligence and Blockchain Guangzhou University Guangdong Key Laboratory of Intelligent Information Processing Shenzhen Key Laboratory of Media Security Shenzhen University Shenzhen Institute of Artificial Intelligence and Robotics for Society School of Computer Science and Cyber Engineering Guangzhou University 

出 版 物：《Science China(Information Sciences)》 (中国科学:信息科学(英文版))

年 卷 期：2023年第66卷第3期

页      面：273-286页

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 0839[工学-网络空间安全] 08[工学] 0835[工学-软件工程] 081201[工学-计算机系统结构] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by National Natural Science Foundation of China (Grant Nos. 62002075, 61872244, 61872099, U19B2022) Guangdong Basic and Applied Basic Research Foundation (Grant No. 2019B151502001) Shenzhen R&D Program (Grant No. JCYJ20200109105008228) 

主　　题：information hiding non-additive steganography steganalysis cost learning image processing 

摘      要：Image steganography is the art and science of secure communication by concealing information within digital images. In recent years, the techniques of steganographic cost learning have developed rapidly. Although the existing methods can learn satisfactory additive costs, the interplay of different pixels’ embedding impacts has not been considered, so the potential of learning may not be fully exploited. To overcome this limitation, in this paper, a reinforcement learning paradigm called Jo Po L(joint policy learning) is proposed to extend the idea of additive cost learning to a non-additive situation. Jo Po L aims to capture the interactions within pixel blocks by defining embedding policies and evaluating contributions of embedding impacts on a block level rather than a pixel level. Then, a policy network is utilized to learn optimal joint embedding policies for pixel blocks through interactions with the environment. Afterwards,these policies can be converted into joint embedding costs for practical message embedding. The structure of the policy network is designed with an effective attention mechanism and incorporated with the domain knowledge derived from traditional non-additive steganographic methods. The environment is responsible for assigning rewards according to the impacts of the sampled joint embedding actions, which are evaluated by the gradient information of a neural network-based steganalyzer. Experimental results show that the proposed non-additive method Jo Po L significantly outperforms the existing additive methods against both feature-based and CNN-based steganalzyers over different payloads.