Multi-Feature Fusion for Malicious PDF Document Detection Based on CNN-BiLSTM-CBAM

doi:10.3969/j.issn.1671-1122.2025.10.009

Abstract

Abstract:

In order to solve the problems that the existing detection methods of malicious PDF documents ignore the semantic relationship between features and are often limited to a single type of feature analysis, this paper proposed a detection scheme, which applied the CNN-BiLSTM-CBAM model and multi-feature fusion to the detection of malicious PDF documents. This method not only integrated the conventional and structural information extracted from static analysis, but also combined the API sequence information captured by dynamic analysis to build a comprehensive multi-dimensional feature set. First, the model used convolutional neural network to extract local features of feature set. Secondly, BiLSTM was used to capture the dependency and context-semantic relationship between features, and convolution block attention module (CBAM) was used to assign different weights to different features to screen out the most distinguishable key features. Finally, Softmax classifier was used to calculate the detection results. The experimental results show that compared with the existing methods, the proposed model shows significant advantages in key performance indicators such as accuracy, recall and F1 score, and effectively improves the detection performance of malicious PDF documents.

Key words: malicious PDF document detection, multi-feature fusion, convolutional block attention module, BiLSTM

CLC Number:

TP309

WANG Youhe, SUN Yi. Multi-Feature Fusion for Malicious PDF Document Detection Based on CNN-BiLSTM-CBAM[J]. Netinfo Security, 2025, 25(10): 1579-1588.

Figures/Tables 13

References 24

[1]	MITRE Corporation. CVE Details Vulnerability Statistics[EB/OL]. (2024-03-22)[2025-05-17]. https://www.cvedetails.com/product/497/Adobe-Acrobat-Reader.html?vendor_id=53.
[2]	YU Min, JIANG Jianguo, LI Gang, et al. A Survey of Research on Malicious Document Detection[J]. Journal of Cyber Security, 2021, 6(3): 54-76.
	喻民, 姜建国, 李罡, 等. 恶意文档检测研究综述[J]. 信息安全学报, 2021, 6(3): 54-76.
[3]	GOPINATH M, SETHURAMAN S C. A Comprehensive Survey on Deep Learning Based Malware Detection Techniques[EB/OL]. (2022-12-21)[2025-05-17]. https://www.sciencedirect.com/science/article/abs/pii/S1574013722000636.
[4]	SRNDIC N, LASKOV P. Detection of Malicious PDF Files Based on Hierarchical Document Structure[C]// ISOC. The 20th Annual Network & Distributed System Security Symposium. San Jose: Citeseer, 2013: 1-16.
[5]	LU Xiaofeng, WANG Fei, JIANG Cheng, et al. A Universal Malicious Documents Static Detection Framework Based on Feature Generalization[EB/OL]. (2021-12-20)[2025-05-17]. https://www.mdpi.com/2076-3417/11/24/12134.
[6]	FALAH A, PAN Lei, HUDA S, et al. Improving Malicious PDF Classifier with Feature Engineering: A Data-Driven Approach[J]. Future Generation Computer Systems, 2021, 115: 314-326. doi: 10.1016/j.future.2020.09.015 URL
[7]	ISSAKHANI M, VICTOR P, TEKEOGLU A, et al. PDF Malware Detection Based on Stacking Learning[C]// Science and Technology Publications. The 8th International Conference on Information Systems Security and Privacy (ICISSP 2022). Heidelberg: Springer, 2022: 562-570.
[8]	AL-HAIJA Q A, ODEH A, QATTOUS H. PDF Malware Detection Based on Optimizable Decision Trees[EB/OL]. (2022-09-30)[2025-05-17]. https://www.mdpi.com/2079-9292/11/19/3142.
[9]	YERIMA S Y, BASHAR A. Explainable Ensemble Learning Based Detection of Evasive Malicious PDF Documents[EB/OL]. (2023-07-20)[2025-05-17]. https://www.mdpi.com/2079-9292/12/14/3148.
[10]	BABAAGBA K O, ADESANYA S O. A Study on the Effect of Feature Selection on Malware Analysis Using Machine Learning[C]// ACM. The 2019 8th International Conference on Educational and Information Technology. New York: ACM, 2019: 51-55.
[11]	YU Yuanzhe, WANG Jinshuang, ZOU Xia. A Malicious PDF Detection Method Based on Feature Agglomeration and Convolutional Neural Network[J]. Information Technology and Network Security, 2021, 40(8): 35-41.
	俞远哲, 王金双, 邹霞. 基于特征集聚和卷积神经网络的恶意PDF文档检测方法[J]. 信息技术与网络安全, 2021, 40(8): 35-41.
[12]	JEONG Y S, WOO J, KANG A R. Malware Detection on Byte Streams of PDF Files Using Convolutional Neural Networks[EB/OL]. (2019-04-03)[2025-05-17]. https://onlinelibrary.wiley.com/doi/10.1155/2019/8485365.
[13]	TOGACAR M, ERGEN B. Processing 2D Barcode Data with Metaheuristic Based CNN Models and Detection of Malicious PDF Files[EB/OL]. (2024-05-09)[2025-05-17]. https://www.sciencedirect.com/science/article/abs/pii/S1568494-624004964.
[14]	JIANG Jianguo, WANG Chenhao, YU Min, et al. NFDD: A Dynamic Malicious Document Detection Method without Manual Feature Dictionary[EB/OL]. (2021-09-09)[2025-05-17]. https://link.springer.com/chapter/10.1007/978-3-030-86130-8_12.
[15]	LEI Jingwei, YI Peng, CHEN Xiang. PDF Document Detection Model Based on Graph Neural Network and Deep Learning[J]. Computer Engineering and Design. 2024, 45(2): 356-366.
	雷靖玮, 伊鹏, 陈祥. 基于图神经网络与深度学习的PDF文档检测模型[J]. 计算机工程与设计, 2024, 45(2): 356-366.
[16]	WANG Wenbo, YI Peng, KOU Taotao, et al. GLDOC: Detection of Implicitly Malicious MS-Office Documents Using Graph Convolutional Networks[J]. Cybersecurity, 2024, 7(1): 48-62. doi: 10.1186/s42400-024-00243-7
[17]	THAKUR P, KANSAL V, RISHIWAL V. Hybrid Deep Learning Approach Based on LSTM and CNN for Malware Detection[J]. Wireless Personal Communications, 2024, 136(3): 1879-1901. doi: 10.1007/s11277-024-11366-y
[18]	MANIRIHO P, MAHMOOD A N, CHOWDHURY M J M. API-MalDetect: Automated Malware Detection Framework for Windows Based on API Calls and Deep Learning Techniques[EB/OL]. (2023-07-22)[2025-05-17]. https://www.sciencedirect.com/science/article/pii/S1084804523001236.
[19]	AFZAL S, ASIM M, JAVED A R, et al. URLdeepDetect: A Deep Learning Approach for Detecting Malicious URLs Using Semantic Vector Models[J]. Journal of Network and Systems Management, 2021, 29(3): 1-27. doi: 10.1007/s10922-020-09571-8
[20]	JIHADO A A, GIRSANG A S. Hybrid Deep Learning Network Intrusion Detection System Based on Convolutional Neural Network and Bidirectional Long Short-Term Memory[J]. Journal of Advances in Information Technology, 2024, 15(2): 219-232. doi: 10.12720/jait.15.2.219-232 URL
[21]	THEKKEKARA J P, YONGCHAREON S, LIESAPUTRA V. An Attention-Based CNN-BiLSTM Model for Depression Detection on Social Media Text[EB/OL]. (2024-03-22)[2025-05-17]. https://www.sciencedirect.com/science/article/pii/S0957417424007000.
[22]	YANG Xiuzhang, PENG Guojun, LUO Yuan, et al. OMRDetector: A Method for Detecting Obfuscated Malicious Requests Based on Deep Learning[J]. Chinese Journal of Computers, 2022, 45(10): 2167-2189.
	杨秀璋, 彭国军, 罗元, 等. OMRDetector:一种基于深度学习的混淆恶意请求检测方法[J]. 计算机学报, 2022, 45(10): 2167-2189.
[23]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional Block Attention Module[C]// Springer. The European Conference on Computer Vision(ECCV 2018). Heidelberg: Springer, 2018: 3-19.
[24]	CIC. PDF Dataset: CIC-Evasive-PDFMal2022[EB/OL]. (2022-02-10)[2025-05-17]. https://www.unb.ca/cic/datasets/PDFMal-2022.html.

特征名称	安全相关性
pdfsize	由于页面大小和内容不同，一般恶意文档较小
metadata size	元数据提供PDF文件信息，可以利用它隐藏嵌入的内容
contains text	文本内容呈现不是恶意PDF文件的目的，因此文件中可能包含较少的文本
header	PDF头部混淆是逃避反病毒扫描的常见方法，恶意文件往往会修改头部格式
obj、endobj	obj对象开始和结束的数量，表示文档中存在对象的个数，正常文档两者相等
stream、endstream	PDF 中二进制数据序列的数量，正常文档两者相等
encrypt	恶意文档倾向于通过加密来隐藏恶意行为
/ObjStm	定义数据流对象，可以隐藏特定的其他对象（文本、脚本、网页链接、图像等）
/JS、/JavaScript	恶意文档往往嵌套JS代码，利用JS代码执行恶意操作，实现堆喷射或解析漏洞等
/JBIG2Decode	指明PDF使用JBIG2压缩，是常见的恶意内容编码过滤器
/launch	执行动作或程序的行为，执行Action的次数与OpenAction字段关联
/EmbeddedFile	恶意文档往往利用嵌入的恶意文件（如Word文档、图片等）执行恶意行为
/XFA	某些PDF文件中包含的XML表单架构，可能存在被攻击者利用的脚本技术
/AA、/Acroform /OpenAction	大部分嵌有JavaScript代码的恶意PDF文件都有自动执行JavaScript代码的行为，其中Acrofrom代表Adobe Acrobat创建的表单，可能被攻击者制作恶意脚本

模型	Precision	Recall	F1-Score	Accuracy	FPR	训练时间/s
RF	97.792%	95.944%	96.859%	96.889%	2.167%	66.175
SVM	94.737%	83.478%	88.752%	89.420%	4.638%	73.394
AdaBoost	97.428%	88.389%	92.689%	93.028%	2.333%	89.728
CNN	98.861%	96.444%	97.638%	97.667%	2.278%	112.327
TextCNN	99.042%	99.087%	99.064%	99.122%	1.049%	125.331
本文模型	99.898%	99.639%	99.768%	99.847%	0.271%	133.172

模型	Precision	Recall	F1-Score	Accuracy	FPR	训练时间/s
CNN	98.861%	96.444%	97.638%	97.667%	2.278%	112.327
CNN-CBAM	99.432%	99.134%	99.283%	99.185%	0.975%	113.244
BiLSTM	98.729%	99.333%	99.030%	99.083%	1.167%	152.692
BiLSTM-CBAM	99.592%	99.491%	99.541%	99.333%	1.085%	154.733
CNN-BiLSTM	99.527%	99.596%	99.561%	99.741%	0.778%	165.038
CNN-BiLSTM-Attention	99.796%	99.694%	99.745%	99.629%	0.543%	148.317
CNN-BiLSTM-CBAM	99.898%	99.639%	99.768%	99.847%	0.271%	133.172

特征类型	Precision	Recall	F1-Score	Accuracy
静态特征	98.387%	96.536%	97.453%	97.333%
API序列	99.243%	98.134%	98.685%	99.085%
静态+API特征	99.898%	99.639%	99.768%	99.847%

方法	类型	Accuracy	Precision	Recall	F1-Score
文献[7]	静态分析	98.69%	98.88%	98.87%	98.87%
文献[14]	动态分析	99.05%	98.48%	99.64%	99.06%
文献[15]	动态分析	99.14%	98.95%	98.87%	98.91%
本文	动静结合分析	99.85%	99.90%	99.64%	99.77%