Medoid Prototype Alignment for Cross-Plant Unknown Attack Detection in Industrial Control Systems

1. Problem Statement

The paper addresses the critical challenge of deploying Intrusion Detection Systems (IDS) across different Industrial Control System (ICS) plants.

• Domain Shift: ICS traffic is highly site-dependent due to differences in sensors, protocols, and control processes. A model trained on one plant (Source) often fails when deployed on another (Target).
• Data Scarcity & Unknown Attacks: Collecting labeled attack data is expensive and risky in production environments. Furthermore, target plants often face "unknown attacks" (attacks not seen during training), making standard supervised learning ineffective.
• Limitations of Existing Methods: Current transfer learning approaches often rely on global feature alignment (aligning all source and target samples). In ICS environments, which contain bursty communication, noise, and structural imbalances, this indiscriminate alignment can obscure local patterns and lead to unstable detection.

2. Methodology: Medoid Prototype Alignment (MPA)

The proposed framework reformulates cross-plant detection as a prototype-guided adaptation problem. Instead of aligning individual data points, it aligns robust structural summaries (prototypes) of the operational regions. The pipeline consists of four main stages:

A. Shared Representation Construction

Since source and target domains may have different feature dimensions ($d_s \neq d_t$), the method first standardizes traffic features independently. It then uses Principal Component Analysis (PCA) to project both domains into a shared, low-dimensional representation space ($Z_s$ and $Z_t$). This creates a uniform basis for comparison.

B. Medoid Prototype Extraction

Rather than using cluster centroids (means), which are sensitive to outliers, the method employs K-Medoids clustering on the PCA-projected data.

• Medoids: These are actual observations from the dataset that minimize the within-cluster dissimilarity.
• Advantage: Medoids represent physically meaningful operating points and are more robust to noise and rare perturbations common in industrial traffic compared to mean-based centers.
• The result is a set of source prototypes ($P_s$) and target prototypes ($P_t$).

C. Prototype-Calibrated Adaptation

The core learning objective involves a representation encoder ($f_\theta$) and a classifier ($g_\phi$). The method defines three loss components:

1. Prototype Alignment Loss ($L_{proto}$): Computes a soft correspondence between target prototypes and source prototypes using a softmax similarity based on Euclidean distance. It minimizes the distance between a target prototype and its weighted source prototypes.
2. Supervised Classification Loss ($L_{sup}$): Ensures the model maintains discriminative power on the labeled source domain.
3. Entropy Regularization ($L_{ent}$): Encourages confident predictions on the unlabeled target domain by minimizing prediction entropy.

The total objective is: $L = L_{sup} + \alpha L_{proto} + \beta L_{ent}$.

3. Key Contributions

1. Novel Problem Formulation: Shifts the paradigm from direct global instance matching to prototype-level transfer, emphasizing the transfer of local structural summaries.
2. Framework Design: Proposes a specific architecture combining PCA compression, K-Medoids extraction, and prototype-calibrated domain adaptation. This design is tailored for heterogeneous, noisy, and label-scarce industrial environments.
3. Empirical Rigor: Redesigns the evaluation to focus on average cross-task performance, directional transfer behavior (asymmetry), and worst-case robustness, providing a more realistic view of deployment reliability than single-task accuracy.

4. Experimental Results

The method was evaluated on two distinct industrial systems: a Natural Gas Control System (G) and a Water Storage Tank System (W). Four cross-domain unknown attack transfer tasks were tested (e.g., DoS attacks from Gas to Water systems).

• Performance Metrics:
  • MPA achieved an average Accuracy of 0.843 and an F1-score of 0.838.
  • It significantly outperformed baselines (ANN, SVM, RF, KNN, NBM). For instance, compared to the strongest baseline (ANN), MPA improved average accuracy by 62.8% and F1-score by 109.4%.
• Robustness:
  • Low Dispersion: MPA showed the lowest coefficient of variation (CV) and the smallest "composite directional gap," indicating consistent performance regardless of the transfer direction (Gas→Water vs. Water→Gas).
  • Worst-Case Stability: Even on the hardest transfer task, MPA maintained a minimum accuracy of 0.81 and F1-score of 0.80, whereas baselines dropped significantly (e.g., ANN dropped to 0.50 accuracy).
• Transfer Asymmetry: The study confirmed that transfer is asymmetric; using the Gas system as a source generally yielded better results than the Water system due to richer control semantics. However, MPA minimized the performance drop in difficult reverse-transfer settings.

5. Significance and Conclusion

The paper demonstrates that Medoid Prototype Alignment is a practical and robust solution for cross-plant intrusion detection.

• Stability over Density: By aligning compact medoid summaries rather than dense sample sets, the method avoids the noise and instability inherent in global alignment for industrial data.
• Deployment Readiness: The high worst-case performance and low variance suggest that MPA is suitable for real-world deployment where reliability under domain shift is critical.
• Future Directions: The authors suggest extending this approach to open-set labeling, online adaptation, and real-time monitoring.

In summary, this work provides a structurally grounded alternative to conventional domain adaptation, proving that leveraging robust local operational structures (via medoids) is superior for handling the heterogeneity and scarcity of labels in industrial cybersecurity.