首页|期刊导航|网络与信息安全学报|基于分层强化学习的图像隐写检测训练方法

基于分层强化学习的图像隐写检测训练方法OA

Hierarchical reinforcement learning-based training method for image steganalysis

中文摘要英文摘要

图像隐写检测是数字取证与多媒体安全领域的重要研究方向,旨在精准区分载体图像与隐写图像.随着自适应隐写算法向低嵌入率、高隐蔽性方向发展,现有基于深度学习的隐写检测模型面临训练策略优化与超参数敏感问题.针对这些问题,提出一种基于分层强化学习的图像隐写检测训练方法.该方法将分层强化学习引入隐写检测器训练过程,通过高低层智能体分别调控回合级与批次级策略,从训练模式和样本权重两个层面对隐写检测器训练进行自适应优化,同时分层智能体通过检测器直接性能指标构建共享的奖励信号,聚焦检测器性能,实现协同优化.在BOSSBase数据集上开展的实验结果表明,所提方法在多种嵌入率场景下均能较好地提升检测性能.所提方法相较XuNet,在0.1~0.5 bpp嵌入率下的平均检测准确率从87.5%提升至90.0%,且适配SRNet与Yedroudj-Net等检测性能较强的隐写检测器.消融实验验证了所提方法的状态设计与分层架构的有效性.

Image steganography detection is a crucial research direction in the field of digital forensics and multimedia security,aiming to accurately distinguish between cover images and steganographic images.As adaptive steganographic algorithms evolve toward low embedding rates and high undetectability,existing deep learning-based steganographic de-tection models face challenges in training strategy optimization and hyperparameter sensitivity.To address these issues,a hierarchical reinforcement learning-based training method for image steganography detection was proposed.In this method,hierarchical reinforcement learning was introduced into the training process of the steganographic detector,where high-level and low-level agents were employed to regulate episode-level and batch-level strategies,respectively.Adaptive optimization of the detector training was achieved from both the training mode and sample weighting perspec-tives.Meanwhile,a shared reward signal was constructed for the hierarchical agents based on the direct performance metrics of the detector,thereby focusing on detector performance and enabling collaborative optimization.Experimental results on the BOSSBase dataset demonstrated that the proposed method consistently improved detection performance under various embedding rate scenarios.Compared with XuNet,the proposed method increased the average detection accuracy from 87.5%to 90.0%under embedding rates ranging from 0.1 to 0.5 bpp,and it was also compatible with more powerful steganographic detectors such as SRNet and Yedroudj-Net.Ablation studies validated the effectiveness of the state space design and the hierarchical architecture in the proposed method.

任德翔;杨荣健;卢晓珍;方黎明

南京航空航天大学,江苏 南京 211106||东南大学移动通信全国重点实验室,江苏 南京 211111南京航空航天大学,江苏 南京 211106南京航空航天大学,江苏 南京 211106||东南大学移动通信全国重点实验室,江苏 南京 211111南京航空航天大学,江苏 南京 211106

信息技术与安全科学

图像隐写检测训练策略优化分层强化学习

image steganography detectiontraining strategy optimizationhierarchical reinforcement learning

《网络与信息安全学报》 2026 (4)

102-114,13

国家自然科学基金资助项目(No.U22B2062,No.62571240,No.U22B2029,No.62272228)江苏省前沿引领技术基础研究重大项目(No.BK20222001)东南大学移动通信全国重点实验室开放研究基金资助课题(No.2026D06)The National Natural Science Foundation of China(No.U22B2062,No.62571240,No.U22B2029,No.62272228),Major Project for Frontier Leading Technology Basic Research of Jiangsu Province(No.BK20222001),The Open Research Fund of the National Mobile Communications Research Laboratory of Southeast University(No.2026D06)

10.11959/j.issn.2096-109x.AQ26055

评论