基于硬件性能计数器的恶意软件检测方法研究

doi:10.3969/j.issn.1671-1122.2026.09.005

摘要/Abstract

摘要：

针对传统恶意软件检测方法在应对加壳、混淆等对抗技术时能力不足的问题，文章提出一种基于硬件性能计数器的恶意软件检测方法。该方法利用Transformer编码器对硬件性能计数器的时序数据进行全局建模，提取表征微架构行为动态演化模式的深度时序特征，进而由XGBoost分类器基于该特征执行最终分类决策，实现时序归纳偏置与结构化归纳偏置的互补。为验证文章所提方法，实验采集了516个恶意样本与500个良性样本在短、中、长3种观测窗口下硬件性能计数器的时序数据进行分析。实验结果表明，该方法在3种窗口下的检测准确率均超过95%，且性能变异系数低于对比方法，展现出良好的检测精度与跨窗口稳定性。上述互补机制有效提升了分类边界的稳健性，为基于微架构行为分析的恶意软件检测提供了一种兼顾表征能力与泛化性能的检测方法。

关键词: 硬件性能计数器, 恶意软件检测, Transformer, XGBoost

Abstract:

To address the limited capability of traditional malware detection methods against evasion techniques such as packing and obfuscation, this paper proposed a malware detection method based on hardware performance counter. The method employed a Transformer encoder to perform global modeling on hardware performance counter time-series data, extracting deep temporal features that capture the dynamic evolution patterns of microarchitectural behaviors. An XGBoost classifier then performed the final classification decision based on these features, achieving a complementary integration of temporal inductive bias and structured inductive bias. Experiments were conducted on hardware performance counter time-series data collected from 516 malicious samples and 500 benign samples under short, medium, and long observation windows respectively to evaluate the proposed method. The results demonstrate that the framework achieves detection accuracy exceeding 95% across all three window configurations, and the performance coefficient of variation is lower than that of the comparison method, exhibiting favorable detection precision and cross-window stability. The above complementary mechanism effectively enhances the robustness of the classification boundary, offering a detection method that balances representational capacity and generalization performance for malware detection based on microarchitectural behavior analysis.

Key words: hardware performance counter, malware detection, Transformer, XGBoost

中图分类号:

TP309

张健, 蔡鸿, 王柏韬. 基于硬件性能计数器的恶意软件检测方法研究[J]. 信息网络安全, 2026, 26(9): 1383-1390.

Zhang Jian, Cai Hong, Wang Botao. Research of malware detection method based on hardware performance counter[J]. Netinfo Security, 2026, 26(9): 1383-1390.

图/表 9

图1

表1

图2

表2

图3

表3

表4

表5

表6

参考文献 22

[1]	Skoudis E, Zeltser L. Malware: fighting malicious code[M]. Upper Saddle River: Prentice Hall Professional, 2004.
[2]	Sikorski M, Honig A. Practical malware analysis: the hands-on guide to dissecting malicious software[M]. San Francisco: No Starch Press, 2012.
[3]	Chandran S, Syam S R, Sankaran S, et al. From static to AI-driven detection: a comprehensive review of obfuscated malware techniques[J]. IEEE Access, 2025(13): 74335-74358.
[4]	Malone C, Zahran M, Karri R. Are hardware performance counters a cost effective way for integrity checking of programs[C]// The 6th ACM Workshop on Scalable Trusted Computing. New York: ACM, 2011: 71-76.
[5]	Raff E, Barker J, Sylvester J, et al. Malware detection by eating a whole EXE[C]// The Workshops of the 32nd AAAI Conference on Artificial Intelligence. Menlo Park: AAAI, 2018: 268-276.
[6]	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)[2026-03-13]. https://onlinelibrary.wiley.com/doi/epdf/10.1155/2019/8485365.
[7]	Ling Xiang, Wu Lingfei, Deng Wei, et al. MalGraph: hierarchical graph neural networks for robust windows malware detection[C]// IEEE INFOCOM 2022-IEEE Conference on Computer Communications. New York: IEEE, 2022: 1998-2007.
[8]	Kolosnjaji B, Zarras A, Webster G, et al. Deep learning for classification of malware system call sequences[C]// The 29th Australasian Joint Conference on Artificial Intelligence (AI 2016). Heidelberg: Springer, 2016: 137-149.
[9]	Gibert D, Mateu C, Planes J, et al. Using convolutional neural networks for classification of malware represented as images[J]. Journal of Computer Virology and Hacking Techniques, 2019, 15(1): 15-28. doi: 10.1007/s11416-018-0323-0
[10]	Wong G W, Huang Yiting, Guo Yingren, et al. Attention-based API locating for malware techniques[J]. IEEE Transactions on Information Forensics and Security, 2024(19): 1199-1212.
[11]	Demme J, Maycock M, Schmitz J, et al. On the feasibility of online malware detection with performance counters[C]// The 40th Annual International Symposium on Computer Architecture (ISCA). New York: ACM, 2013: 559-570.
[12]	Tang A, Sethumadhavan S, Stolfo S J. Unsupervised anomaly-based malware detection using hardware features[C]// International Workshop on Recent Advances in Intrusion Detection (RAID 2014). Heidelberg: Springer, 2014: 109-129.
[13]	Ozsoy M, Donovick C, Gorelik I, et al. Malware-aware processors: a framework for efficient online malware detection[C]// 2015 IEEE 21st International Symposium on High Performance Computer Architecture (HPCA). New York: IEEE, 2015: 651-661.
[14]	严飞, 彭慧容, 何凡, 等. HBROP: 基于硬件性能计数器的函数级ROP检测[J]. 武汉大学学报(理学版), 2017, 63(2): 109-116.
[15]	邢凌凯, 张健. 基于HPC的虚拟化平台异常检测技术研究与实现[J]. 信息网络安全, 2023, 23(10): 64-69.
[16]	Wang Xueyang, Karri R. Reusing hardware performance counters to detect and identify kernel control-flow modifying rootkits[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2016, 35(3): 485-498. doi: 10.1109/TCAD.2015.2474374 URL
[17]	Sayadi H, Makrani H M, Pudukotai D S M, et al. 2SMaRT: a two-stage machine learning-based approach for run-time specialized hardware-assisted malware detection[C]// 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE). New York: IEEE, 2019: 728-733.
[18]	Elnaggar R, Servadei L, Mathur S, et al. Accurate and robust malware detection: running XGBoost on runtime data from performance counters[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2022, 41(7): 2066-2079. doi: 10.1109/TCAD.2021.3102007 URL
[19]	Kuruvila A P, Meng Xingyu, Kundu S, et al. Explainable machine learning for intrusion detection via hardware performance counters[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2022, 41(11): 4952-4964. doi: 10.1109/TCAD.2022.3149745 URL
[20]	He Zhangying, Miari T, Makrani H M, et al. When machine learning meets hardware cybersecurity: delving into accurate zero-day malware detection[C]// 2021 22nd International Symposium on Quality Electronic Design (ISQED). New York: IEEE, 2021: 85-90.
[21]	Zhou Boyou, Gupta A, Jahanshahi R, et al. Hardware performance counters can detect malware: myth or fact?[C]// The 2018 on Asia Conference on Computer and Communications Security. New York: ACM, 2018: 457-468.
[22]	Weaver V M. Self-monitoring overhead of the Linux perf_event performance counter interface[C]// 2015 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). New York: IEEE, 2015: 102-111.

序号	事件名称	含义
1	instructions	执行指令总数
2	cpu-cycles	CPU运行周期数
3	cache-misses	末级缓存未命中次数
4	L1-dcache-loads	L1数据缓存加载次数
5	branch-misses	分支预测失败次数
6	branch-instructions	分支指令执行次数
7	iTLB-load-misses	指令TLB加载未命中次数
8	LLC-stores	末级缓存存储次数

数据集	样本总数/个	恶意样本数/个	正常样本数/个
训练集	711	361	350
验证集	152	77	75
测试集	153	78	75
合计	1016	516	500

采样窗口配置	模型	准确率	精确率	召回率	F1分数
10 s_100 ms	RF	95.42%	91.76%	100%	0.9571
	XGB	94.12%	89.66%	100%	0.9455
	DT	92.81%	88.51%	98.72%	0.9333
	SVM	81.05%	72.90%	100%	0.8432
20 s_200 ms	RF	94.77%	92.68%	97.44%	0.9500
	XGB	95.42%	92.77%	98.72%	0.9565
	DT	94.77%	91.67%	98.72%	0.9506
	SVM	82.35%	74.29%	100%	0.8525
30 s_300 ms	RF	92.81%	91.36%	94.87%	0.9308
	XGB	94.77%	94.87%	94.87%	0.9487
	DT	94.12%	93.67%	94.87%	0.9427
	SVM	88.89%	82.80%	98.72%	0.9006

采样配置	模型	准确率	精确率	召回率	F1分数
10 s_100 ms	LSTM	92.16%	87.50%	98.72%	0.9277
	BiLSTM	93.46%	91.46%	96.15%	0.9375
	BiLSTM_Attn	91.50%	86.52%	98.72%	0.9222
	TCN	92.81%	87.64%	100%	0.9341
	Transformer	92.81%	90.36%	96.15%	0.9317
20 s_200 ms	LSTM	92.81%	87.64%	100%	0.9341
	BiLSTM	93.46%	88.64%	100%	0.9398
	BiLSTM_Attn	94.12%	89.66%	100%	0.9455
	TCN	92.81%	87.64%	100%	0.9341
	Transformer	94.77%	93.75%	96.15%	0.9494
30 s_300 ms	LSTM	94.12%	89.66%	100%	0.9455
	BiLSTM	91.50%	85.71%	100%	0.9231
	BiLSTM_Attn	94.77%	92.68%	97.44%	0.9500
	TCN	93.46%	88.64%	100%	0.9398
	Transformer	95.10%	92.77%	98.72%	0.9530

方法	10 s_100 ms		20 s_200 ms		30 s_300 ms		平均值
方法	准确率	F1分数	准确率	F1分数	准确率	F1分数	准确率	F1分数
文献[11]方法	92.16%	0.9125	92.81%	0.9183	93.46%	0.9210	92.81%	0.9173
文献[13]方法	93.46%	0.9289	94.12%	0.9367	94.77%	0.9412	94.12%	0.9356
文献[14]方法	94.77%	0.9458	94.77%	0.9462	94.12%	0.9437	94.55%	0.9452
本文方法	95.98%	0.9612	95.72%	0.9581	95.37%	0.9533	95.69%	0.9575