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

Jiacen Xu, Chenang Li, Yu Zheng, Zhou Li

TL;DR

ENTENTE is a new federated learning (FL) framework designed to train graph-based intrusion detection systems across distributed data silos. By using synthetic reference graphs and adaptive weighting, it allows organizations to build powerful security models that detect sophisticated attacks (like APTs) while keeping sensitive network logs strictly on-premise and compliant with privacy regulations like GDPR.

What Problem Does This Paper Address?

Modern cyber-attacks, particularly Advanced Persistent Threats (APTs), are increasingly sophisticated, often moving laterally across a network over long periods. To catch them, security researchers use Graph-based Network Intrusion Detection Systems (GNIDS), which model network logs as massive graphs to spot abnormal interactions.

The catch? These models usually require a centralized "data lake" to work effectively. However, in the era of GDPR and PDPA, centralizing sensitive network logs across different regions or companies is often illegal or operationally impossible. While Federated Learning (FL)—where models are trained locally and only "summaries" are shared—seems like a solution, it typically fails on network data. This is because network graphs are highly "non-IID" (one office's network looks nothing like another's) and are easily subverted by attackers who can "poison" the global model during the training process.

Methods

ENTENTE introduces several novel techniques to make federated graph learning practical for security:

  • Reference Graph Synthesis & Sketching: To handle the diversity of different networks, the central server generates a synthetic "reference graph" (using a Barabási-Albert model). Clients use Weisfeiler-Lehman (WL) sketching to compare their local network structure to this reference, allowing the server to initialize client weights more intelligently.
  • Adaptive Contribution Scaling (ACS): Instead of giving every client equal say, ENTENTE dynamically adjusts a client's weight in each training round based on how much their local model diverges from the global one. This helps the model converge faster even when data is messy and imbalanced.
  • Dynamic Clipping (Robustness): By combining ACS with norm-bounding, ENTENTE implements a "dynamic clipping" mechanism. It automatically throttles updates that look like outliers, effectively neutralizing model poisoning attacks without the accuracy loss associated with traditional static clipping.
  • 1-hop Graph Augmentation: Since silos often miss connections to each other, ENTENTE "harvests" cross-client edges from local firewall and proxy logs for free, filling in the gaps in the global graph structure without requiring direct data sharing between clients.

Datasets

The authors evaluated ENTENTE on three large-scale, real-world network log datasets:

  1. LANL (Los Alamos National Laboratory): A massive dataset covering 58 days and over 1 billion events.
  2. OpTC (Operationally Transparent Cyber): A DARPA-funded dataset focused on high-fidelity endpoint telemetry.
  3. Pivoting: A specialized dataset for detecting lateral movement (pivoting) within a network.

Key Insights and Ideas

  • The "Heterogeneity" Hurdle: Standard FL assumes data is somewhat uniform. In network security, one silo might have 100 nodes while another has 10,000. ENTENTE proves that sharing just a tiny bit of metadata (like the total node count) to build a reference graph can bridge this gap.
  • FL Can Beat Centralized Training: Surprisingly, ENTENTE sometimes outperformed "Non-FL" (centralized) models. This is because the federated approach, when combined with smart weighting, can better handle the "long-tail" distribution of rare malicious events that often get washed out in a single massive dataset.
  • Robustness is Mandatory: In a federated setting, you must assume at least one client is compromised. ENTENTE’s ability to bound the Attack Success Rate (SR) to low levels (e.g., <10%) while maintaining high detection accuracy is a significant step toward making FL trustworthy for mission-critical security.

Key Results

  • Superior Accuracy: On the OpTC dataset, ENTENTE achieved an Average Precision (AP) of 74-84%, outperforming SOTA federated baselines by a wide margin.
  • High Detection Rates: For link prediction (identifying suspicious connections), ENTENTE reached an AUC of over 0.9 in most test cases across LANL and Pivoting.
  • Scalability: The system proved it could scale to 50+ clients. While it adds some computational overhead (a few seconds per iteration), it drastically reduces data transfer—sending only 1.94MB of model weights instead of 65GB of raw logs.
  • Poisoning Resistance: Under adaptive poisoning attacks, ENTENTE remained stable, whereas baseline FL methods often suffered from "gradient explosion" (NaN errors) or saw their detection rates plummet.

Limitations and Future Work

  • Label Granularity: The authors note that detection results on LANL are still challenging because the ground-truth labels are "coarse-grained," making it hard to distinguish the exact start and end of a malicious campaign.
  • Differential Privacy (DP) Trade-offs: While ENTENTE supports DP to further protect against privacy leaks (like membership inference attacks), the authors acknowledge that adding enough noise for strong privacy still causes a noticeable drop in detection accuracy.
  • Real-world Distribution: The study used MBM clustering to simulate silos. Future work could test ENTENTE in a truly distributed environment across different physical organizations to see how latency and real-world network jitter affect performance.

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},
}

Loading related content...

Figure