Siamese Neural Networks for User Identity Linkage Through Web Browsing.

Area of Science:

• Computer Science
• Data Science
• Cybersecurity

Background:

• User identity linkage (UIL) is crucial for research and industry, enabling a unified digital user representation across diverse online services.
• Existing UIL methods often focus on social networks or sparse location data, limiting their applicability to general web-browsing behaviors.
• Distinguishing users based on web-browsing patterns is essential for applications like identifying malicious users or children online.

Purpose of the Study:

• To extend User Identity Linkage (UIL) to general web-browsing behaviors, moving beyond social networks.
• To develop a robust model capable of handling the inherent challenges of UIL, including imbalanced datasets.
• To accurately link online identities by analyzing deep feature representations of user web-browsing activities.

Main Methods:

• Proposed a Siamese Neural Network (NN) architecture-based UIL (SAUIL) model to learn and compare high-level feature representations of web-browsing behaviors.
• Introduced a cost-sensitive SAUIL (C-SAUIL) model to address the issue of imbalanced datasets in UIL by assigning higher misclassification costs to minority classes.
• Utilized deep neural networks for feature extraction and comparison of user web-browsing patterns.

Main Results:

• The proposed SAUIL and C-SAUIL models demonstrated robustness and good performance on large-scale, real-world web-browsing datasets.
• The C-SAUIL model effectively addressed the challenge of imbalanced UIL data, improving classification accuracy for minority classes.
• The models successfully linked online identities based on web-browsing behaviors, outperforming previous approaches in general scenarios.

Conclusions:

• The developed Siamese NN-based models (SAUIL and C-SAUIL) offer a powerful solution for User Identity Linkage (UIL) using web-browsing data.
• Addressing data imbalance is critical for effective UIL, and the cost-sensitive approach significantly enhances model performance.
• This research provides a scalable and effective method for UIL across heterogeneous internet services, with implications for cybersecurity and user behavior analysis.