MirGuard: Towards a Robust Provenance-based Intrusion Detection System Against Graph Manipulation Attacks

Anyuan Sang, Lu Zhou, Li Yang, Junbo Jia, Huipeng Yang, Pengbin Feng, Jianfeng MaarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces MirGuard, a novel intrusion detection system that uses provenance graphs to identify malicious activity. MirGuard is designed to be robust against "graph manipulation attacks" where attackers subtly alter these graphs to hide their tracks, achieving high detection accuracy with minimal performance degradation.

What Problem Does This Paper Address?

Modern cybersecurity relies heavily on detecting anomalies in system behavior. Provenance graphs, which map out the relationships between system events and entities (like processes, files, and network connections), are powerful tools for this. They provide a rich context that can reveal complex attack chains. However, existing intrusion detection systems (PIDSes) that use these graphs are vulnerable to "graph manipulation attacks." Attackers can strategically modify the provenance graph – for example, by adding "fake" benign-looking connections or altering node features – to make malicious activities appear normal, thus evading detection. While this threat has been recognized, existing PIDSes lack robust defenses against these sophisticated manipulation techniques, limiting their real-world applicability.

Methods

MirGuard tackles this challenge with a two-pronged approach:

  • Logic-Aware Graph Augmentation: Instead of randomly perturbing the graph (which can create unrealistic scenarios), MirGuard uses "Logic-Aware Noise Injection" (LNI). This method generates multiple "views" of the provenance graph by introducing perturbations (adding/removing nodes/edges/features) that are constrained by the inherent logic and semantics of provenance data. For example, it won't add an edge directly between a file and a network connection if that doesn't make sense in the system's context. This ensures the augmented graphs are still semantically valid, mimicking more realistic adversarial scenarios.
  • Logic-Preserving Contrastive Learning: These semantically valid graph views are then fed into a contrastive learning framework. The goal here is to train the model to learn representations that are:
  • Invariant to benign, logic-preserving transformations (meaning it recognizes the same underlying behavior even if slightly altered).
  • Sensitive to adversarial inconsistencies (meaning it can detect when something is "off" due to malicious manipulation). This contrastive learning approach helps the model distinguish between genuine system behavior and manipulated data, even when the manipulations are subtle.

For anomaly detection itself, MirGuard uses a KMeans clustering-based approach on the learned representations, which is efficient and effective.

Datasets

MirGuard was evaluated on several widely used provenance datasets:

  • DARPA TC THEIA, CADETS, TRACE: These datasets are used for node-level detection.
  • StreamSpot, Unicorn Wget: These datasets are used for graph-level detection.

The paper notes that these datasets vary in their scenarios and detection granularity, providing a comprehensive evaluation.

Key Insights and Ideas

  • Semantics over Superficial Structure: The core idea is that attackers can easily manipulate the superficial structure of a graph, but it's much harder to maintain the underlying causal semantics. MirGuard's "logic-aware" approach focuses on preserving these semantics, making it resilient to structural tampering.
  • Realistic Adversarial Simulation: By constraining augmentations to domain-specific logic, MirGuard creates more realistic adversarial examples than random noise injection, leading to more effective robustness training.
  • Contrastive Learning for Robustness: The multi-view contrastive learning framework is crucial. It forces the model to learn representations that are stable against benign changes but sensitive to malicious ones, effectively building robustness into the core of the detection model.
  • Efficiency Matters: The paper emphasizes that robustness shouldn't come at the cost of performance. MirGuard aims to maintain high detection accuracy and efficiency comparable to existing methods.

Key Results

  • Superior Robustness: MirGuard significantly outperforms state-of-the-art detectors (Threatrace, MAGIC, FLASH) against various graph manipulation attacks (GSPA, GFPA, CGPA, SPA, FPA), even at high attack rates (e.g., 50%). For instance, under CGPA-50%, MirGuard achieved an F1-score of 0.872 and an AUC of 0.975, far exceeding baselines.
  • High Detection Accuracy: MirGuard maintains near-perfect detection performance (e.g., F1-scores of 0.99 and AUCs close to 1.0) across multiple datasets, even under attack conditions.
  • Minimal Performance Degradation: Under graph manipulation attacks, MirGuard shows minimal degradation in its F1-score (less than 10% drop) and AUC compared to baselines.
  • Efficiency: MirGuard's training and inference times are comparable to or better than some baseline methods, demonstrating that robustness is achieved without sacrificing efficiency.

Limitations and Future Work

  • Trusted Computing Base (TCB) Assumption: The system assumes the integrity of the underlying operating system and auditing framework. Hardware trojans or side-channel attacks that bypass these mechanisms are not addressed.
  • Focus on Graph Manipulation: While MirGuard excels against graph manipulation, its effectiveness against other types of sophisticated attacks (e.g., zero-day exploits not easily represented in provenance graphs) might require further investigation.
  • Future Work: The authors suggest exploring different augmentation strategies and detection methods to further enhance robustness and efficiency. The paper also highlights the potential for MirGuard's logic-preserving contrastive learning approach to be applied to other graph-based security tasks.

Summary updated

Abstract

Learning-based Provenance-based Intrusion Detection Systems (PIDSes) have become essential tools for anomaly detection in host systems due to their ability to capture rich contextual and structural information, as well as their potential to detect unknown attacks. However, recent studies have shown that these systems are vulnerable to graph manipulation attacks, where attackers manipulate the graph structure to evade detection. While some previous approaches have discussed this type of attack, none have fully addressed it with a robust detection solution, limiting the practical applicability of PIDSes. To address this challenge, we propose MirGuard, a robust anomaly detection framework that combines logic-aware multi-view augmentation with contrastive representation learning. Rather than applying arbitrary structural perturbations, MirGuard introduces Logic-Aware Noise Injection (LNI) to generate semantically valid graph views, ensuring that all augmentations preserve the underlying causal semantics of the provenance data. These views are then used in a Logic-Preserving Contrastive Learning framework, which encourages the model to learn representations that are invariant to benign transformations but sensitive to adversarial inconsistencies. Comprehensive evaluations on multiple provenance datasets demonstrate that MirGuard significantly outperforms state-of-the-art detectors in robustness against various graph manipulation attacks without sacrificing detection performance and efficiency. Our work represents the first targeted study to enhance PIDS against such adversarial threats, providing a robust and effective solution to modern cybersecurity challenges.
BibTeX
@misc{sang2025mirguard,
  author = {Anyuan Sang and Lu Zhou and Li Yang and Junbo Jia and Huipeng Yang and Pengbin Feng and Jianfeng Ma},
  title = {{MirGuard: Towards a Robust Provenance-based Intrusion Detection System Against Graph Manipulation Attacks}},
  year = {2025},
  eprint = {2508.10639},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.10639},
}
Figure