基于频率感知与义原增强的文本防御编码  

Textual defense encoding based on frequency aware and sememe enhance

作　　者：罗浩岚 刘万平[1] 王宝娟 黄东 LUO Hao-lan;LIU Wan-ping;WANG Bao-juan;HUANG Dong(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China;Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education,Guizhou University,Guiyang 550025,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054 [2]贵州大学现代制造技术教育部重点实验室,贵州贵阳550025

出　　处：《计算机工程与设计》2025年第3期749-755,共7页Computer Engineering and Design

基　　金：重庆市自然科学基金项目(cstc2021jcyj-msxmX0594);重庆理工大学研究生教育高质量发展行动计划资助成果基金项目(gzlcx20233260)。

摘　　要：针对文本防御编码未考虑训练样本中词频的影响,同义词集缺乏囊括性且存在一定噪声的问题,提出一种基于频率感知与义原增强的编码训练方法。引入样本单词频率,利用编码器区分为样本中的低频词与非低频词,分别训练其鲁棒性;替换词集采用义原增强后的样本数据,能够有效扩充现有词集;编码算法能使样本有效训练确保模型原始准确率。在常见数据集上的实验结果表明,编码训练下的模型分类准确率优于之前防御方法,分别在TextCNN与LSTM上降低模型平均误差到3.6%与4.2%。Aiming at the problems that text defense encoding does not consider the influence of word frequency in the training sample,and the synonym sets lack of inclusiveness and have some noise,an encoding training method based on frequency awareness and sememe enhancement was proposed.Introducing the sample word frequency,the encoder was used to distinguish the low-frequency words and non low-frequency words in the sample and its robustness was trained respectively.The replacement word set effectively expanded the existing word set by using the sample data after sememe enhancement.The samples were effectively trained using the encoding algorithm to ensure the original accuracy of the model.Experimental results on common datasets show that the classification accuracy of the model under encoding training is better than that of the previous defense methods,and the average error of the model is reduced to 3.6% and 4.2% on TextCNN and LSTM.

关 键 词：文本防御编码 深度神经网络 文本分类 同义词替换攻击 频率感知 义原增强 文本对抗样本 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]