基于域适应的电磁泄漏还原图像中文文本识别

吕志强; 于超; 李海洋; 张宁

引用本文：

吕志强,于超,李海洋,张宁.基于域适应的电磁泄漏还原图像中文文本识别[J].信息安全学报,2026,11(2):258-272 [点击复制]
LV Zhiqiang,YU Chao,LI Haiyang,ZHANG Ning.Chinese Text Recognition in Electromagnetic Emission Reconstructed Images Based on Domain Adaptive[J].Journal of Cyber Security,2026,11(2):258-272 [点击复制]

本文已被：浏览 13次下载 8次	码上扫一扫！
基于域适应的电磁泄漏还原图像中文文本识别
吕志强^1,2, 于超^1,2, 李海洋^1,2, 张宁^1,2
0 字体:加大+\|默认\|缩小-
(1.中国科学院信息工程研究所第四研究室北京中国 100093;2.中国科学院大学网络空间安全学院北京中国 100093)

摘要:

计算机显示系统会在信息的传输和显示过程中产生电磁泄漏,利用TEMPEST技术(Transient Electromagnetic Pulse Emanation Surveillance Technology),可以很容易地将辐射的电磁信息截获,在通过电磁泄漏途径获取的视频图像中,图像中的文字往往含有十分重要的信息,也是我们更为关注的内容,因此对于从电磁泄漏途径得到的图像,其文字区域的识别是一项至关重要的工作,然而通过接收机接收的电磁泄漏的视频信号信噪比很低,然而通过接收机接收的电磁泄漏的视频信号信噪比很低,这使得还原的图像难以进行有效的文本识别。现有的针对低信噪比中文文本图像的文字识别工作非常少。在本文中,我们提出了一种基于域适应思想的CRNN (Convolutional Recurrent Neural Network)文字识别模型。该模型用电磁泄漏环境下采集的无标注文本图像作为目标域数据,正常的带标注文本图像作为源域数据,将卷积神经网络(Convolutional Neural Network,CNN)结合上域判别模块(Domain Discrimination Module,DDM),然后采用半监督学习的训练方式使得卷积神经网络最终提取的特征层是带随机噪声的目标域数据集和正常的源域数据集的公共特征,由于是两者的公共特征,因此也就最小化各种随机噪声带来的影响,并且可以最大化地利用这些鲁棒的公共特征来进行后续的字符分类。提升了真实噪声环境条件下的文字识别准确率。本文模型在电磁泄漏还原实景下的公开数据集RCTW-17、CASIA-10上进行了测试,评价指标为精确率(Precision)和归一化平均编辑距离(Normalized Average Edit Distance,NAED),相比于主流的识别模型,基于域适应的CRNN对于电磁泄漏还原的文本图像的精确率和归一化平均编辑距离有了明显的提升。

关键词: 电磁泄漏文本识别域适应半监督学习神经网络

DOI：10.19363/J.cnki.cn10-1380/tn.2026.03.16

投稿时间：2020-12-24修订日期：2021-03-03

基金项目:本课题得到国家重点研发计划(No.2018YFF01014303)资助。

Chinese Text Recognition in Electromagnetic Emission Reconstructed Images Based on Domain Adaptive

LV Zhiqiang^1,2, YU Chao^1,2, LI Haiyang^1,2, ZHANG Ning^1,2

(1.The 4th Laboratory, Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China;2.School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100093, China)

Abstract:

Electromagnetic emission exists in the process of information transmission and display in computer display system. Using TEMPEST technology (Transient Electrical Pulse Analysis Surveillance Technology), radiated electromagnetic information can be easily intercepted. In video images obtained through electromagnetic leakage, the text in the image often contains very important information, which is also the focus of our attention. Therefore, for images obtained through electromagnetic leakage, the recognition of the text area is a crucial task. However, the signal-to-noise ratio of the emitted video signal received by the receiver is very low, and it makes the restored image difficult for effective text recognition. There are few text recognition methods for Chinese text images with low signal-to-noise ratio. In this paper, We propose a CRNN (Convolutional Recurrent Neural Network) text recognition model based on domain adaptation, which uses the unlabeled text images collected in the electromagnetic emission environment as the target domain data, and uses the normal labeled text images as the source domain data. The model combines the Convolutional Neural Network (CNN) with the Domain Discrimination Module(DDM), and then then the semi supervised learning training method is adopted to make the final feature layer extracted by the convolutional neural network be the common features of the target domain dataset with random noise and the normal source domain dataset. As they are common features of both, the impact of various random noise is minimized, and these robust common features can be maximized for subsequent character classification. which improves the accuracy of text recognition in images emitted from target computer. This model was tested on publicly available datasets RCTW-17 and CASIA-10k in the context of electromagnetic leakage restoration, and the evaluation indicators were Precision and Normalized Average Edit Distance (NAED). Compared with mainstream recognition models, The domain adaptation based CRNN has significantly improved the accuracy and normalized average editing distance of text images restored by electromagnetic leakage.

Key words: electromagnetic emission text recognition domain adaptive few-shot learning neural network