BERT-Based Approaches to Identifying Malicious URLs.

Area of Science:

• Cybersecurity
• Machine Learning
• Natural Language Processing

Background:

• Malicious Uniform Resource Locators (URLs) pose significant threats in cyberattacks, including phishing and malware distribution.
• Accurate detection of malicious URLs is crucial for robust cybersecurity measures.
• Previous research utilized deep learning models for URL analysis, often segmenting strings into tokens for classification.

Purpose of the Study:

• To develop and evaluate a novel approach for malicious URL detection using a transformer-based model.
• To leverage the self-attention mechanism of BERT for improved understanding of token correlations within URL strings.
• To assess the model's performance across diverse datasets, including specialized domains like IoT and DoH.

Main Methods:

• A Bidirectional Encoder Representation from Transformers (BERT) model was employed for tokenizing URL strings.
• The self-attention mechanism within BERT was utilized to capture inter-token relationships.
• A classification layer was integrated to distinguish between malicious and benign URLs.
• The model was evaluated on three public datasets (Kaggle, GitHub, ISCX 2016) and two domain-specific datasets (IoT, DoH).

Main Results:

• The proposed BERT-based system achieved high accuracy rates of 98.78%, 96.71%, and 99.98% on the three primary public datasets.
• The model demonstrated versatility and effectiveness when tested on Internet of Things (IoT) and Domain Name System over HTTPS (DoH) datasets.
• The self-attention mechanism proved effective in enhancing the model's comprehension of URL structures.

Conclusions:

• The developed BERT-based model offers a highly accurate and versatile solution for malicious URL detection.
• This approach significantly advances the capabilities of automated systems in identifying and mitigating cyber threats.
• The model's strong performance across various datasets highlights its potential for real-world cybersecurity applications.