一种基于深度学习的移动端隐写方法

来源:用户上传      作者:廖鑫 黎懿熠 欧阳军林 周江盟 戴湘桃 秦拯

　　摘要：隐写是隐蔽通信的主流方法之一，而移动端则是当下最常用的通信设备，二者的结合研究具有较高的实际意义.近年来，基于深度学习的隐写方法得到快速发展，然而在性能提升的同时，各类网络结构向着更复杂、庞大的方向演变，逐渐脱离以隐蔽通信为核心的实际应用场景，实用性较低.针对这一现象，本文提出一种适用于移动端的轻量级图像隐写方法.对网络整体进行轻量化设计，结合深度可分离卷积降低模型计算量，在精度和速度之间取得较好的折中平衡.以生成对抗网络的思想，将编码器、解码器和判别器构成的整体模型纳入对抗训练中，使子网络在迭代对弈中实现螺旋式上升发展.为应对真实环境下的各类挑战，模型被落地部署于移动设备上进行真机实验.在移动端，精简后的模型性能会出现小幅下降.对此，在方法中引入BCHm错码以确保正确提取信息.实验结果表明，该移动端隐写方法生成图像质量好，且具有较高的响应速度，能满足现代社会中人们对便捷性的高要求.值得注意的是，该方法的所有计算工作均可在移动端独立完成，不需要通过网络请求服务器，能避免网络窃听攻击.
　　关键词：隐写;深度学习;生成对抗网络;移动端;轻量级
　　中图分类号：TP309文献标志码：A
　　A Mobile Steganography Method Based on Deep Learning
　　LIAO Xin LI Yiyi OUYANG Junlin ZHOU Jiangmeng DAI Xiangtao QIN Zheng1
　　（1. College of Computer Science and Electronic Engineering，Hunan University，Changsha 41008 China;
　　2. School of Computer Science and Engineering，Hunan University of Science and Technology，Xiangtan 41120 China;
　　3. School of Physics and Electronics，Central South University，Changsha 41008 China;
　　4. Great Wall Information Co.，Ltd，Changsha 410199，China）
　　Abstract：Steganography is one of the main methods for covert communication，while mobile phones are the most commonly used communication devices. The combination of the two has high practical significance. In recentyears，steganography has developed rapidly with deep learning technologies. To improve the performance，networks evolve towards a more complex and large style，which gradually deviates from the real world scenarios with covert communication as the core，resulting in low practicability. For convenience and efficiency，a lightweight image steganography method is proposed for mobile phone. The network structure is designed in a light style，with depthwise separable convolutions utilized to reduce useless parameters and keeping a balance between accuracy and speed. Based on generative adversarial networks，the proposed method consists of a generator，a decoder，and a discriminator，which are trained together defiantly and finally advance in a spiral upward trend. To deal with various challenges in the real world，the model is deployed on mobile phones for tests. The networks used on smartphones are pruned，which indicates performance degradation. To ameliorate this problem and enhance decoding accuracy，BCH correcting codes are used in the method. The results show that the method can generate high-quality images with high speed，which meets the convenience requirements in today's world. Besides，it's worth noting that the method works without online requests. All the embedding and extracting tasks can be done by phone itself，which means this scheme is immune to eavesdropping attacks.

nlc202208291746

转载注明来源:https://www.xzbu.com/4/view-15438676.htm