Robust Botnet Detection Against Reinforcement-Learning Evasion: Transferability, Operational Cost and a Grounded Analyst Assistant

Author

Netrang V Davey, Rishab P K, Dr. G. Paavai Anand

Abstract

 Tree-based classifiers are a common choice for flow-level botnet detection because they are accurate on benchmark traffic and cheap to run. Their weakness is that an attacker who can adjust a few flow statistics, without breaking the underlying connection, can push malicious flows across the decision boundary. The DReLAB benchmark packages exactly this scenario: it pairs original botnet flows with versions rewritten by deep reinforcement learning agents. We use it to ask three questions that earlier work on this dataset left open. First, when a detector is hardened with adversarial flows from some botnet families, does the protection carry over to a family it has never seen? Second, what does hardening cost on ordinary traffic, measured as false alarms? Third, do the features that the attacker edits coincide with the features the model actually relies on? A Random Forest trained on 171 thousand clean flows reaches 97.89% recall at 0.70% false-positive rate, yet detects only 25.51% of the adversarial flows. Mixing 10% adversarial samples into training lifts adversarial recall to 95.84% while the false-positive rate moves from 0.71% to 0.83% and clean F1 stays at 98.6%. In leave-one-family-out experiments the hardened model gains between 36 and 68 recall points on five unseen families, but only 11.7 points on Murlo. SHAP analysis shows that the two most heavily perturbed features rank first and third in the model's own importance ordering. We also describe a small LangGraph assistant whose language model can only report numbers returned by seven deterministic tools wrapped around the trained model, which lets an analyst question individual verdicts without risk of fabricated statistics.

Keywords

Botnet detection; adversarial evasion; reinforcement learning; adversarial training; transferability; false-positive rate; SHAP; network flow features; LangGraph; security operations.

Full Text:

Download Paper PDF

References

[1] G. Apruzzese, M. Andreolini, M. Marchetti, A. Venturi, and M. Colajanni, “Deep reinforcement adversarial learning against botnet evasion attacks,” IEEE Transactions on Network and Service Management, vol. 17, no. 4, pp. 1975–1987, 2020.

[2] A. Venturi, G. Apruzzese, M. Andreolini, M. Colajanni, and M. Marchetti, “DReLAB – Deep REinforcement Learning Adversarial Botnet: A benchmark dataset for adversarial attacks against botnet intrusion detection systems,” Data in Brief, vol. 34, 106631, 2021.

[3] S. García, M. Grill, J. Stiborek, and A. Zunino, “An empirical comparison of botnet detection methods,” Computers & Security, vol. 45, pp. 100–123, 2014.

[4] I. Sharafaldin, A. H. Lashkari, and A. A. Ghorbani, “Toward generating a new intrusion detection dataset and intrusion traffic characterization,” in Proc. 4th International Conference on Information Systems Security and Privacy (ICISSP), pp. 108–116, 2018.

[5] M. Ring, S. Wunderlich, D. Scheuring, D. Landes, and A. Hotho, “A survey of network-based intrusion detection data sets,” Computers & Security, vol. 86, pp. 147–167, 2019.

[6] I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” in Proc. International Conference on Learning Representations (ICLR), 2015.

[7] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Advances in Neural Information Processing Systems, vol. 30, pp. 4765–4774, 2017.

[8] S. M. Lundberg et al., “From local explanations to global understanding with explainable AI for trees,” Nature Machine Intelligence, vol. 2, pp. 56–67, 2020.

[9] S. Yao et al., “ReAct: Synergizing reasoning and acting in language models,” in Proc. International Conference on Learning Representations (ICLR), 2023.

[10] LangChain Inc., “LangGraph documentation,” 2024. https://langchain-ai.github.io/langgraph/

[11] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.

[12] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785–794, 2016.

Share your valuable work from Social Media Buttons