项目名称: 基于深度学习的隐写分析新方法研究
项目编号: No.61303262
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 董晶
作者单位: 中国科学院自动化研究所
项目金额: 27万元
中文摘要: 本课题拟研究基于深度学习的结构化隐写分析新方法。现有的隐写方法都是以启发式的特征设计、浅层机器学习为主,得到有效隐写分析特征后,多采用监督学习的方法得到隐写分类检测结果。这类方法在特定训练样本库能够获得较好的检测性能。然而,面向当前互联网海量、多源、异质数据的隐写分析时,其检测性能将急剧下降。本课题旨在研究图像邻域相关性分析的基础上,通过深度学习的方式获取最优的隐写特征表达,并通过引入鉴别信息,设计多层次网络结构的深度学习算法,提高隐写分析算法在处理海量多源图像的准确性、鲁棒性和泛化性,以期满足信息安全领域的重大应用需求。
中文关键词: 信息隐藏;隐写分析;深度学习;增强CNN;迁移学习
英文摘要: This project is aiming at image steganlaysis based on deap learning. Current steganlaysis technicques are mostly based on heuristic feature extraction and surpervised learning. These methods can achive good detection results on specific and limited testing samples. However, facing the big data era, surpervised learning based steganlysis methods can't meet their demands, unwell trained classifier usually fails on new samples and can not updated online. Our research is focus on steganalysis using deep learning. The steganlaysis feature will be extracted on image neighborhood structure. We also will design the deep learning algorithms and proper networks for classification. The authentication information will be used to increase the robustness and generality for detection.
英文关键词: Steganalysis;Deep learning;CNN;Transfer learning;