基于DQN生成对抗样本的JavaScript恶意代码检测模型  

A JAVASCRIPT MALICIOUS CODE DETECTION MODEL BASED ON DQN GENERATION OF ADVERSARIAL SAMPLES

作　　者：苏庆[1] 温炜亮 林佳锐 黄剑锋[1] 谢国波[1] Su Qing;Wen Weiliang;Lin Jiarui;Huang Jianfeng;Xie Guobo(School of Computers,Guangdong University of Technology,Guangzhou 510006,Guangdong,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《计算机应用与软件》2025年第3期332-340,共9页Computer Applications and Software

基　　金：国家自然科学基金项目(618002072);广东省自然科学基金项目(2018A030313389);广东省高等教育教学改革项目(SJJG20191216)。

摘　　要：针对基于深度学习的JavaScript恶意代码检测模型抗攻击能力较弱的问题,提出一个基于DQN(Deep Q-Learning Network)生成对抗样本的JavaScript恶意代码检测组合模型DQN-CNN。利用CNN对数据集进行训练,得到初始判别器origin_CNN。将DQN作为生成器,两者组成DQN-origin_CNN对抗模型进行训练。在训练过程中DQN通过代码混淆动作,生成origin_CNN的对抗样本。接着将对抗样本加入数据集,对origin_CNN持续进行迭代训练,获得最终判别器retrain_CNN。实验结果表明,retrain_CNN与DQN组成新的对抗模型DQN-retrain_CNN生成对抗样本成功率显著下降,从45.7%下降为21.5%,证明最终生成的判别器retrain_CNN的抗攻击能力得到了显著提升。To address the problem that JavaScript malicious code detection models based on deep learning are weak against attacks,a combined model DQN-CNN for JavaScript malicious code detection based on DQN generation of adversarial samples is proposed.The initial discriminator origin_CNN was obtained by training the dataset with CNN.The DQN was used as a generator and the two formed a DQN-origin_CNN adversarial model for training.During the training process,DQN generated the adversarial samples of origin_CNN by code obfuscation actions.The adversarial samples were added to the dataset,and the origin_CNN was continuously trained iteratively to obtain the final discriminator retrain_CNN.The experimental results show that the success rate of generating adversarial samples for the new adversarial model DQN-retrain_CNN composed of retrain_CNN and DQN decreases significantly,from 45.7%to 21.5%,proving that the final generated discriminator retrain_CNN has significantly improved its resistance to attacks.

关 键 词：深度强化学习 代码混淆 灰度图 JAVASCRIPT代码 对抗攻击 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]