Explainable few-shot learning with dynamic prototypes for distributed fiber-optic intrusion detection.

Area of Science:

• Cybersecurity
• Machine Learning
• Signal Processing

Background:

• Intrusion detection systems (IDS) are crucial for infrastructure security.
• Distributed optical fiber vibration sensing (DVS) offers perimeter security but struggles with limited data and model interpretability.
• Existing deep learning models often lack transparency, hindering trust.

Purpose of the Study:

• To develop an explainable intrusion detection framework for DVS systems.
• To address challenges of data scarcity (few-shot learning) and model interpretability.
• To enhance the reliability and trustworthiness of perimeter security solutions.

Main Methods:

• Proposed an explainable dual-branch feature fusion dynamic class center prototypical network (DBFF-DC-ProtoNet).
• Utilized a lightweight dual-branch 1-D ResNet for temporal and time-frequency feature extraction.
• Integrated a dynamic class center update strategy, novel loss function, and an explainability module (Proto-CAM, case-based reasoning).

Main Results:

• Achieved high accuracy (97.22% and 98.33%) in 5-shot settings on benchmark datasets.
• Demonstrated effective fusion of temporal and time-frequency features for discriminative prototypes.
• Showcased fine-grained signal attribution and intuitive case retrieval for enhanced explainability.

Conclusions:

• DBFF-DC-ProtoNet successfully bridges few-shot learning with interpretability for DVS intrusion detection.
• The proposed model offers a practical and trustworthy solution for securing perimeters with limited labeled data.
• The framework enhances both the performance and transparency of intrusion detection systems.