Anomaly detection system for computing power networks based on behavioral attention

doi:10.3969/j.issn.1671-1122.2026.07.008

Abstract

Abstract:

As a new-generation information infrastructure, the efficient and stable operation of computing power networks serves as a critical cornerstone for ensuring the security of the national digital economy and the resilience of key information infrastructure. With the continuous growth in computing power demand and the increasing scale and complexity of tasks, accurate prediction of task demands is not only a prerequisite for optimal allocation of computing resources but also a key means of identifying abnormal resource usage and potential attack behaviors. However, existing methods often struggle to capture local sudden fluctuations and complex dynamic patterns across multiple task types when processing long-sequence task demands, leading to reduced resource scheduling efficiency and issues such as missed detections and false alarms for anomalous demands. To address these challenges, this paper designed and implemented an anomalous task demand detection system for computing power networks. The system introduced a behavior-enhanced attention mechanism that, by integrating multi-scale statistical measures with sliding window analysis, dynamically captures task attributes and temporal behavioral features, thereby significantly improving the identification and prediction of sudden patterns. Through precise sensing of demand surge behaviors in the network, the system helped detect potential precursors of anomalous traffic attacks in advance, providing decision support for secure scheduling of computing resources and system protection. Experimental results on the Alibaba Cluster Traces dataset demonstrate that the behavior-enhanced attention mechanism integrated into the system achieves optimal performance in terms of both mean squared error (MSE) and mean absolute error (MAE), significantly outperforming existing methods in perception accuracy for complex task fluctuations. In summary, the proposed system not only improves the accuracy and robustness of task demand prediction in computing power networks but also enhances the perception of potential risks.

Key words: computing power networks, demand forecast, behavior enhanced attention, anomaly detection

CLC Number:

TP309

Qian Xiaosheng, Wang Liying, Zhao Yu, Liu Guangya, Wang Cheng. Anomaly detection system for computing power networks based on behavioral attention[J]. Netinfo Security, 2026, 26(7): 1101-1114.

Figures/Tables 17

References 28

[1]	张宏科, 于成晓, 权伟, 等. 融算网络体系基础研究[J]. 电子学报, 2022, 50(12):2928-2934. doi: 10.12263/DZXB.20221140
[2]	邵子豪, 王志浩, 周晓茂, 等. 算电协同研究综述:架构、关键技术与展望[J]. 通信学报, 2025, 46 (10) : 287-308.
[3]	王璐, 张健浩, 王廷, 等. 面向云网融合的细粒度多接入边缘计算架构[J]. 计算机研究与发展, 2021, 58(6):1275-1290.
[4]	段晓东. 2023年算力网络趋势产业链携手构建共同体[J]. 通信世界, 2023(2):26-27.
[5]	任晓旭, 谭靖超, 邓辉, 等. 基于端边云超融合的算力网络架构[J]. 计算机应用, 2022(z1):195-200.
[6]	Zhang Ke, Xiao Xiong, PENG Zhenwen, et al. MCA: model compromise attacks against federated computing power networks[C]// The 2023 International Conference on Electronics, Computers and Communication Technology. New York: ACM, 2023: 215-220.
[7]	莫益军. 算力网络场景需求及算网融合调度机制探讨[J]. 信息通信技术, 2022, 16(2):34-39.
[8]	Sun Hongju, Qiu Qin, XU Jiawei, et al. Computing force network security: risks, protection requirements and scenarios[C]//2024 ITU Kaleidoscope:Innovation and Digital Transformation for a Sustainable World (ITU K). New York: IEEE, 2024: 1-8.
[9]	Liu Yong, Hu Tengge, Zhang Haoran, et al. iTransformer: inverted transformers are effective for time series forecasting[C/OL]// The 12th International Conference on Learning Representations, 2024: 1. https://openreview.net/forum?id=JePfAI8fah.
[10]	Zhang Yunhao, Yan Junchi. Crossformer: transformer utilizing cross-dimension dependency for multivariate time series forecasting[C/OL]// The 11th International Conference on Learning Representations, 2023: 1. https://openreview.net/pdf?id=vSVLM2j9eie.
[11]	Liu Yong, Wu Haixu, Wang Jianmin, et al. Non-stationary transformers: Exploring the stationarity in time series forecasting[C]// Advances in Neural Information Processing Systems, 2022, 35: 9881-9893.
[12]	Du Dazhao, Su Bing, Wei Zhewei. Preformer: predictive transformer with multi-scale segment-wise correlations for long-term time series forecasting[C]// 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). New York: IEEE, 2023: 1-5.
[13]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]// Advances in Neural Information Processing Systems. Red Hook: Curran Associates, 2017: 5998-6008.
[14]	Lin Yang, Koprinska I, Rana M. SpringNet: transformer and spring DTW for time series forecasting[C]// International Conference on Neural Information Processing. Heidelberg: Springer, 2020: 616-628.
[15]	Liu Shizhan, Yu Hang, Liao Cong, et al. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting[C/OL]// International Conference on Learning Representations, 2022: 1. https://openreview.net/pdf?id=0EXmFzUn5I.
[16]	Zhou Haoyi, Zhang Shanghang, Peng Jieqi, et al. Informer: beyond efficient transformer for long sequence time-series forecasting[C]// The AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2021: 11106-11115.
[17]	Madhusudhanan K, Burchert J, Born S, et al. U-net inspired transformer architecture for far horizon time series forecasting[C]// Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Heidelberg: Springer, 2022: 36-52.
[18]	Nie Yuqi, Nguyen N H, Sinthong P, et al. A time series is worth 64 words: long-term forecasting with transformers[C/OL]// The Eleventh International Conference on Learning Representations, 2023: 1. https://openreview.net/pdf?id=Jbdc0vTOcol.
[19]	Lim B, Arık S Ö, Loeff N, et al. Temporal fusion transformers for interpretable multi-horizon time series forecasting[J]. International Journal of Forecasting, 2021, 37(4): 1748-1764. doi: 10.1016/j.ijforecast.2021.03.012 URL
[20]	Kang Wangcheng, McAuley J. Self-attentive sequential recommendation[C]// 2018 IEEE International Conference on Data Mining (ICDM). New York: IEEE, 2018: 197-206.
[21]	Sun Fei, Liu Jun, Wu Jian, et al. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer[C]// The 28th ACM International Conference on Information and Knowledge Management. New York: ACM, 2019: 1441-1450.
[22]	Wu Liwei, Li Shuqing, Hsieh C J, et al. SSE-PT: sequential recommendation via personalized transformer[C]// The 14th ACM Conference on Recommender Systems. New York: ACM, 2020: 328-337.
[23]	Kumar S, Zhang Xikun, Leskovec J. Predicting dynamic embedding trajectory in temporal interaction networks[C]// The 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2019: 1269-1278.
[24]	Li Jiacheng, Wang Yujie, McAuley J. Time interval aware self-attention for sequential recommendation[C]// The 13th International Conference on Web Search and Data Mining. New York: ACM, 2020: 322-330.
[25]	Ye Wenwen, Wang Shuaiqiang, Chen Xu, et al. Time matters: sequential recommendation with complex temporal information[C]// The 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2020: 1459-1468.
[26]	Liu Zhiwei, Fan Ziwei, Wang Yu, et al. Augmenting sequential recommendation with pseudo-prior items via reversely pre-training transformer[C]// The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2021: 1608-1612.
[27]	Zhou Zhou, Shojafar M, Alazab M, et al. IECL: an intelligent energy consumption model for cloud manufacturing[J]. IEEE Transactions on Industrial Informatics, 2022, 18(12): 8967-8976. doi: 10.1109/TII.2022.3165085 URL
[28]	Cui Yangguang, Cao Kun, Zhou Junlong, et al. Optimizing training efficiency and cost of hierarchical federated learning in heterogeneous mobile-edge cloud computing[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2022, 42(5): 1518-1531. doi: 10.1109/TCAD.2022.3205551 URL

名称	配置情况
操作系统	Windows 11
CPU	12th Gen Intel(R) Core(TM) i5-12400 2.50 GHz
GPU	NVIDIA GeForce RTX3060
CUDA	CUDA12.0
Cudnn	cudnn11.6.55
开发工具	PyCharm
开发语言	Python3.9
深度学习框架	PyTorch1.13

数据集	特征数/个	时间间隔/h	样本数/个
Worker	4	1	1224
TensorFlow	4	1	1224
PyTorchWorker	4	1	1224

数据集	时刻	本文方法		iTransformer		Crossformer		PatchTST		NS-Trans		FEDformer
数据集	时刻	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
PyTorchWorker	6	0.782	0.597	0.794	0.594	0.822	0.608	0.915	0.624	0.823	0.608	0.834	0.607
	12	0.795	0.586	0.800	0.587	0.849	0.626	0.894	0.625	0.829	0.622	0.812	0.618
	24	0.817	0.63	0.795	0.587	0.846	0.615	0.895	0.634	0.975	0.705	0.853	0.627
	48	0.862	0.627	0.887	0.628	0.917	0.636	0.92	0.648	1.014	0.721	0.907	0.651
Worker	6	1.050	0.761	1.051	0.693	1.070	0.758	1.096	0.763	1.064	0.78	1.083	0.768
	12	1.109	0.766	1.070	0.775	1.098	0.775	1.131	0.776	1.129	0.788	1.104	0.775
	24	1.086	0.791	1.087	0.778	1.148	0.774	1.175	0.802	1.087	0.777	1.109	0.789
	48	1.231	0.827	1.327	0.842	1.423	0.865	1.340	0.851	1.286	0.858	1.278	0.845
TensorFlow	6	1.262	0.813	1.297	0.838	1.277	0.816	1.354	0.853	1.284	0.845	1.265	0.814
	12	1.230	0.816	1.268	0.831	1.238	0.817	1.318	0.840	1.260	0.821	1.234	0.825
	24	1.198	0.813	1.200	0.82	1.214	0.798	1.254	0.832	1.264	0.851	1.202	0.818
	48	1.130	0.808	1.148	0.808	1.153	0.778	1.176	0.818	1.182	0.802	1.170	0.821

数据集	时刻	Transformer		+MBR		+MCL		+MBR & MCL
数据集	时刻	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
PyTorchWorker	6	0.914	0.605	0.825	0.617	0.817	0.608	0.782	0.597
	12	0.882	0.623	0.842	0.627	0.844	0.623	0.795	0.586
	24	0.939	0.640	0.849	0.634	0.903	0.656	0.817	0.630
	48	0.949	0.637	0.938	0.668	0.905	0.631	0.862	0.627
Worker	6	1.084	0.776	1.058	0.763	1.058	0.769	1.050	0.761
	12	1.148	0.794	1.113	0.773	1.126	0.778	1.109	0.766
	24	1.136	0.793	1.115	0.806	1.107	0.792	1.086	0.791
	48	1.245	0.831	1.234	0.830	1.256	0.843	1.231	0.827
Tensorflow	6	1.286	0.832	1.272	0.828	1.264	0.816	1.262	0.813
	12	1.264	0.821	1.230	0.818	1.249	0.819	1.230	0.816
	24	1.205	0.817	1.213	0.815	1.202	0.816	1.198	0.813
	48	1.157	0.812	1.146	0.813	1.154	0.817	1.130	0.808

数据集	预测长度/时间步	无行为增强注意力模块捕获率	有行为增强注意力模块捕获率
Worker	48	10.71%	25.00%
TensorFlow	48	2.28%	5.28%
PytorchWorker	48	2.31%	5.31%