Introduction
Internet of Things (IoT) devices are increasingly prevalent and notoriously vulnerable to cyberattacks. Developing robust Intrusion Detection Systems (IDS) for IoT environments is hampered by a critical lack of labeled data. While traditional network intrusion detection (NID) benefits from decades of data, IoT intrusion detection (IID) operates in a different domain with distinct protocols, hardware, and data formats. This disparity presents a Heterogeneous Domain Adaptation (HDA) challenge: transferring knowledge from a data-rich source domain to a data-scarce target domain with non-overlapping feature spaces.
The Geometric Graph Alignment (GGA) Framework
To address the HDA problem, the authors propose Geometric Graph Alignment (GGA). This approach shifts the focus from statistical distribution alignment to the intrinsic geometric relationships between attack categories. The core insight is that the relative "signature shape" of attacks—such as the relationship between a Denial of Service (DoS) and a Probing attack—remains largely invariant across different network architectures.
The GGA framework employs a hierarchical alignment strategy:
- Shape Keeping: An adversarial discriminator ensures the global structure of attack categories in the common feature subspace is consistent between domains.
- Rotation and Symmetry Correction: Specific loss functions ($L_R$ and $L_{CP}$) prevent spatial misalignments like rotation or mirror symmetry that could occur during feature projection.
- Vertex-level Semantic Preservation: This ensures that specific attack types (e.g., "Backdoor") maintain their semantic meaning across domains by aligning their respective centroids.
Pseudo-Label Election (PLE)
A major risk in domain adaptation is "negative transfer," where noisy or incorrect labels from the target domain degrade model performance. The authors introduce the Pseudo-Label Election (PLE) mechanism to mitigate this. PLE acts as a conservative quality filter for unlabeled IoT data.
An unlabeled instance is only assigned a pseudo-label if three independent "voters" agree: the shared classifier's prediction, the instance's geometric proximity to attack centroids, and the labels of its nearest neighbors. This mechanism effectively filters out instances near ambiguous decision boundaries or those that are geometrically inconsistent, ensuring high-quality data for model refinement.
Experimental Evaluation
The GGA approach was evaluated using five benchmark datasets, including NSL-KDD for source domains and UNSW-BOTIOT for target IoT domains. The model was tested under extreme data scarcity, such as a 1:100 ratio of labeled to unlabeled target instances.
Key Results: * Accuracy: GGA outperformed nine state-of-the-art baselines by an average of 4.2%. Its performance remained robust even as labeled data became increasingly scarce. * Efficiency: GGA is highly suitable for resource-constrained IoT environments. It demonstrated a training speed 31 times faster than deep adaptation networks like DDAC and a 15.79% faster inference speed compared to graph-learning methods like WCGN. * Ablation Studies: Statistical tests confirmed that each component of the hierarchical alignment (shape keeping, rotation avoidance, etc.) contributes significantly to the overall performance boost.
Conclusion
The Geometric Graph Alignment approach provides a powerful solution for IoT intrusion detection by leveraging the geometric consistency of cyberattacks across heterogeneous domains. By combining hierarchical graph alignment with a robust Pseudo-Label Election mechanism, GGA achieves superior accuracy and efficiency. Future research will explore "Open Set" domain adaptation to handle novel attack categories that do not exist in the source training data.
Summary updated