Tor (The onion router) is an anonymous communication network based on multi-layer encryption and distributed routing technology, which is widely used for privacy protection. However, its high anonymity also renders the network a breeding ground for Darknet activities, posing severe threats to national security and social stability. Due to the diversity of node functions and network complexity, effective node classification has become a critical research topic. This paper proposes a Tor node classification method utilizing Dynamic Self-Attention Temporal Graph Neural Networks(DySAT). This approach analyzes the historical association graph of Tor relay nodes and employs a spatiotemporal dual-attention mechanism to simultaneously capture node performance and security indicators. Experiments validate the effectiveness of the proposed method by selecting high-quality nodes and benchmarking performance against normal circuits. Compared with Tor’s default circuit construction algorithm, the probability of selecting malicious nodes is reduced from 6.2% to 1.8%, and the average latency drops from 0.478 s to 0.389 s. Consequently, this method provides a new technical approach for Darknet governance, vulnerable node identification, and mitigation, helping to enhance network security capabilities and contain illegal activities on the Darknet.
实验3 网络泛化能力对比实验。为评估模型在面对网络变化时的泛化能力,实验对比了传统GCN和本文提出的DySAT算法的表现(图6)。实验结果表明,传统GCN由于其直推式学习机制的限制,准确率随时间的下降而显著衰减。在模型训练20 d后,针对当前Tor网络的优质节点筛选,准确率已降至60.7%。相比之下,本文提出的DySAT算法展示了更强的泛化能力。经过20 d 的Tor真实网络的变化,DySAT模型在优质中继节点识别上的准确率仍达到88.8%,与模型训练初期的结果相比,仅下降7.0%。实验结果验证了动态图神经网络在处理节点动态演化时的强适应性。
DingledineR, MathewsonN, SyversonP.Tor: The second-generation onion router [C]//Proceedings of the 13th USENIX Security Symposium, 2004.
[6]
LanH L, LiF J, WangQ, et al.Anonymous communication and darknet Space comprehensive governance [J].Journal of Frontiers of Computer Science and Technology, 2022, 16(11): 2430-2455.
DarirH, BorisovN, DullerudG.DiProber: Using dual probing to estimate tor relay capacities in underloaded networks [PP/OL].V1.arxiv(2022-11-30)[2025-10-13].
[13]
CangialosiF, LevinD, SpringN.Ting: Measuring and exploiting latencies between all tor nodes [C]//Proceedings of the 2015 Internet Measurement Conference, 2015: 289-302.
[14]
SherrM, BlazeM, LooB T.Scalable link-based relay selection for anonymous routing [M]//Privacy enhancing technologies, 2009: 73-93.
[15]
SanatiniaA, NoubirG.Honey Onions: A framework for characterizing and identifying misbehaving Tor HSDirs [C]//2016 IEEE Conference on Communications and Network Security (CNS), 2016: 127-135.
[16]
DahalS, LeeJ, KangJ, et al.Analysis on end-to-end node selection probability in Tor network [C]//2015 International Conference on Information Networking (ICOIN).Cambodia: IEEE, 2015: 46-50.
[17]
WuD, SiS, WangH, et al.Social influence aware sybil detection in social networks [C]//2017 IEEE/CIC International Conference on Communications in China (ICCC), 2017: 1-4.
[18]
WinterP, EnsafiR, LoesingK, et al.Identifying and characterizing sybils in the tor network [C]//USENIX Security Symposium, 2016: 1169-1185.
[19]
MoF.Research on Tor network transmission performance optimization methods [D].Chengdu: University of Electronic Science and Technology of China, 2024.
[20]
莫凡.Tor网络传输性能优化方法研究[D].成都: 电子科技大学, 2024.
[21]
LüM, ZhuY F, LinW.Dynamic routing algorithm based on bandwidth of anonymous network [J].Journal on Communications, 2019(5): 591-596.
WangX, FangB X, LiuP P, et al.Measuring and analyzing node families in the Tor anonymous communication network [J].Journal on Communications, 2015(2): 80-87.
SankarA, WuY, GouL, et al.DySAT: Deep neural representation learning on dynamic graphs via self-attention networks [C]//Proceedings of the 13th International Conference on Web Search and Data Mining, 2020: 519-527.
[26]
WangJ Z, KongL W, HuangZ C, et al.Survey of graph neural network [J].Computer Engineering, 2021, 47(4): 1-12.
WuS, SunF, ZhangW, et al.Graph neural networks in recommender systems: A survey [J].ACM Comput Surv, 2023, 55(5): 1-37.
[31]
LiX.Design and implementation of high quality nodes selection scheme for Tor network [D].Chengdu: University of Electronic Science and Technology of China, 2023.
[32]
李鑫.Tor网络高质量节点选择方案的设计与实现[D].成都: 电子科技大学, 2023.
[33]
SanatiniaA, NoubirG.Honey Onions: A framework for characterizing and identifying misbehaving Tor HSDirs[C]//2016 IEEE Conference on Communications and Network Security (CNS), 2016: 127-135.
[34]
DahalS, LeeJ, KangJ, et al.Analysis on end-to-end node selection probability in Tor network [C]//2015 International Conference on Information Networking (ICOIN), 2015: 46-50.
[35]
LinY.Research on Tor communication performance optimization method based on circuit delay estimation [D].Chengdu: University of Electronic Science and Technology of China, 2022.