A Biological Immunity-Based Neuro Prototype for Few-Shot Anomaly Detection with Character Embedding

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Anomaly detection is crucial for identifying various issues like intrusions and equipment failures.
• Limited annotated data and trusted labels in practical scenarios hinder detection performance.

Purpose of the Study:

• To propose a novel few-shot anomaly detection network for text data, inspired by biological immune systems.
• To address the challenge of low-resource anomaly detection in text-based systems.

Main Methods:

• A character-level representation extraction and Word2Vec embedding approach.
• A meta-learning phase utilizing a dynamic prototype with encoder, routing, and relation modules.
• A dynamic routing algorithm to assign weights to support set samples for improved prototype generation.

Main Results:

• The proposed anomaly detection prototype outperformed state-of-the-art few-shot techniques, achieving 1.3%-4.48% higher accuracy and 0.18%-4.55% higher recall.
• Effective detection was maintained with significantly reduced training samples (5-10).
• Ablation studies confirmed the dynamic routing algorithm's contribution to more accurate prototypes.

Conclusions:

• The biologically inspired anomaly detection network offers superior performance in few-shot text anomaly detection.
• The dynamic routing mechanism is key to creating robust and accurate anomaly detection prototypes.
• This approach effectively mitigates the impact of limited labeled data in anomaly detection tasks.