Entente: Cross-silo Intrusion Detection on Network Log Graphs with Federated Learning

Jiacen Xu, Chenang Li, Yu Zheng, Zhou Li

Introduction

Modern cyber-attacks, particularly Advanced Persistent Threats (APTs), are increasingly sophisticated, often moving laterally across a network over long periods. To detect these threats, security researchers utilize Graph-based Network Intrusion Detection Systems (GNIDS), which model network logs as massive graphs to identify abnormal interactions. However, training these models typically requires a centralized "data lake," which is increasingly problematic under privacy regulations like GDPR and PDPA. Centralizing sensitive network logs across different regions or organizations is often legally or operationally prohibitive.

Overview of distributed network log collection and graph construction across silos, illustrating the challenge of cross-silo lateral movement by an attacker.

While Federated Learning (FL)—where models are trained locally and only model updates are shared—offers a potential solution, it often struggles with network data. Network graphs are inherently "non-IID" (one office's network structure differs vastly from another's) and are vulnerable to model poisoning attacks where a compromised client subverts the global model.

The ENTENTE Framework

ENTENTE is a novel federated learning framework designed to train graph-based intrusion detection systems across distributed data silos while maintaining privacy and robustness. The system coordinates local training at the client level with global aggregation at a parameter server through a multi-step pipeline.

The ENTENTE architecture, detailing the 10-step process from local log processing to global model aggregation, including adaptive scaling and robustness mechanisms.

Key Methodological Innovations

  • Reference Graph Synthesis & Sketching: To address network heterogeneity, the central server generates a synthetic "reference graph." Clients use Weisfeiler-Lehman (WL) sketching to compare their local network structure to this reference, enabling the server to initialize client weights more intelligently.
  • Adaptive Contribution Scaling (ACS): Rather than uniform weighting, ENTENTE dynamically adjusts a client's weight in each training round based on the divergence of their local model from the global one. This accelerates convergence in the presence of imbalanced data.
  • Dynamic Clipping for Robustness: By combining ACS with norm-bounding, ENTENTE implements a dynamic clipping mechanism. It automatically throttles updates that appear as outliers, neutralizing model poisoning attacks without the accuracy degradation typical of static clipping.
  • 1-hop Graph Augmentation: To account for missing cross-silo connections, ENTENTE "harvests" edges from local firewall and proxy logs, filling gaps in the global graph structure without requiring direct data sharing.

Experimental Evaluation

ENTENTE was evaluated on three large-scale, real-world network log datasets: LANL (over 1 billion events), OpTC (high-fidelity endpoint telemetry), and Pivoting (focused on lateral movement).

Key Results

  • Superior Accuracy: On the OpTC dataset, ENTENTE achieved an Average Precision (AP) of 74-84%, significantly outperforming state-of-the-art federated baselines. For link prediction, it reached an AUC of over 0.9 in most test cases.
  • Efficiency and Scalability: The system scales effectively to 50+ clients. It drastically reduces data transfer requirements, sending only 1.94MB of model weights instead of 65GB of raw logs.
  • Poisoning Resistance: ENTENTE maintained stability under adaptive poisoning attacks, bounding the Attack Success Rate (SR) to low levels (e.g., <10%) while baseline methods often suffered from gradient explosion or severe accuracy loss.

Discussion and Future Work

ENTENTE demonstrates that federated learning can occasionally outperform centralized training. This is attributed to the framework's ability to handle the "long-tail" distribution of rare malicious events that might otherwise be obscured in a single massive dataset.

However, challenges remain. Ground-truth labels in datasets like LANL are often "coarse-grained," making precise detection difficult. Furthermore, while ENTENTE supports Differential Privacy (DP), the inherent trade-off between strong privacy noise and detection accuracy remains a hurdle for future research. Future work will focus on testing ENTENTE in truly distributed environments to assess the impact of real-world network latency and jitter.

Summary updated

BibTeX
@inproceedings{xu2026entente,
  author = {Jiacen Xu and Chenang Li and Yu Zheng and Zhou Li},
  title = {{Entente: Cross-silo Intrusion Detection on Network Log Graphs with Federated Learning}},
  booktitle = {NDSS},
  year = {2026},
}
Figure