RPM-Net Reciprocal Point MLP Network for Unknown Network Security Threat Detection

1. Problem Statement

The paper addresses the critical challenge of detecting unknown network security threats in multi-class, imbalanced environments. Current network intrusion detection systems (NIDS) face three primary limitations:

• Open Set Recognition (OSR) Gap: Traditional methods often assume a closed set (all test data belongs to training classes). In reality, attackers use novel techniques (zero-day exploits, new ransomware) that do not fit known patterns.
• Class Imbalance: Real-world network traffic has highly skewed distributions (e.g., benign traffic vastly outnumbers specific attack types, and some attacks are rare). Most existing multi-class OSR methods assume balanced distributions, leading to poor generalization.
• Lack of Interpretability & "Non-Class" Modeling: Existing approaches often rely on simplistic binary classification (normal vs. attack) or lack mechanisms to explicitly model the "unknown" space. They struggle to define boundaries where novel threats can be safely identified without requiring samples of those unknown threats during training.

2. Methodology: RPM-Net Framework

The authors propose RPM-Net, a novel framework designed to learn "non-class" representations for known categories and create geometric boundaries for unknown threats. The architecture consists of four key components:

A. Feature Extractor

A Multi-Layer Perceptron (MLP) with ReLU activations and Dropout regularization extracts feature embeddings ($z$) from raw network flow data.

B. Reciprocal Point Mechanism

Instead of learning a center point for a class (what a class is), RPM-Net learns a Reciprocal Point ($P_k$) for each known class $k$, representing what a class is not.

• Logic: The distance between an embedding $z$ and its reciprocal point $P_k$ is calculated using a combination of normalized Euclidean distance and cosine similarity.
• Classification: A sample is assigned to class $k$ if it is "far" from the reciprocal point of other classes but "close" to the reciprocal point of its own class (conceptually, maximizing the distance to the "non-class" space).

C. Adversarial Margin Constraints

To handle unknown threats, the model enforces a margin constraint ($R_k$) around each reciprocal point.

• Mechanism: Known class samples are constrained to remain within a specific distance ($R_k$) from their reciprocal point.
• Result: This pushes known classes toward the periphery of the feature space, leaving a central "open space" where unknown threats naturally reside. This creates a geometric boundary that separates known patterns from unknown anomalies without needing unknown samples during training.

D. Fisher Discriminant Regularization (RPM-Net++)

To further enhance performance, the authors introduce RPM-Net++, which adds Fisher Discriminant Regularization to the loss function.

• Goal: Maximize the ratio of inter-class scatter (separation between different known classes) to intra-class scatter (compactness within a single class).
• Effect: This ensures that known classes are tightly clustered and well-separated, making the "open space" for unknowns more distinct and improving robustness against class imbalance.

E. Training Objective

The total loss function combines three components:
$$L_{total} = \alpha L_{CE} + \lambda L_{margin} + \beta L_{Fisher}$$
Where $L_{CE}$ is cross-entropy based on reciprocal point distances, $L_{margin}$ enforces the boundary constraints, and $L_{Fisher}$ optimizes class separability.

F. Inference

• Classification: Assigns the label of the reciprocal point with the maximum distance.
• Unknown Detection: If the maximum distance to any reciprocal point is below a threshold $\tau$, the sample is flagged as an unknown threat.

3. Key Contributions

1. Reciprocal Point Mechanism: Introduces a novel way to represent "non-class" spaces for multi-class open set recognition, enabling effective differentiation of known attacks.
2. Adaptive Boundary Construction: Uses adversarial margin constraints to create a feature space where known classes are pushed to the periphery, naturally accommodating unknown threats in the center.
3. Fisher Regularization (RPM-Net++): Enhances intra-class compactness and inter-class separability, specifically addressing the challenges of multi-class imbalanced environments.
4. Unified Training Strategy: The framework learns an open-set recognition feature space using only known attack data, eliminating the need for unknown class samples during the training phase.

4. Experimental Results

The model was evaluated on two standard datasets: CICIDS2017 and UNSW-NB15, comparing against Baseline, ODIN, OCN, and EVM.

• CICIDS2017 Performance:

  • Known Class Classification: Achieved a macro F1-score of 0.9987.
  • Unknown Detection: Achieved an AUPR-OUT (Area Under Precision-Recall Curve for Out-of-Distribution) of 0.6523, significantly outperforming the next best method (EVM at 0.2974).
  • RPM-Net++: Further improved AUROC to 0.9735 and AUPR-OUT to 0.6711.

• UNSW-NB15 Performance:

  • Known Class Classification: F1-score of 0.7950.
  • Unknown Detection: AUPR-OUT of 0.8555, outperforming all baselines.
  • RPM-Net++: Improved AUPR-OUT to 0.8664.

• Ablation Study: The addition of Fisher regularization (RPM-Net++) consistently improved AUROC and AUPR-OUT across both datasets, confirming the synergy between reciprocal points, margin constraints, and discriminative regularization.

5. Significance and Impact

• Practical Applicability: RPM-Net addresses the real-world constraint of class imbalance and the inability to collect data for every potential future attack. It provides a viable solution for early warning systems against zero-day threats.
• Geometric Interpretability: By visualizing the feature space where known classes are pushed to the edges and unknowns occupy the center, the model offers a clear geometric interpretation of how unknown threats are detected.
• State-of-the-Art Performance: The results demonstrate that RPM-Net significantly outperforms existing open-set recognition methods in both accuracy for known attacks and sensitivity to unknown threats, making it a strong candidate for deployment in modern cybersecurity infrastructure.