Ensemble techniques for detecting profile cloning attacks in online social networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Social Network Analysis

Background:

• Online social networks (OSNs) face increasing identity deception challenges.
• AI-generated content and large language models (LLMs) complicate traditional detection methods.
• Detecting sophisticated fake profiles, including cloned and AI-generated ones, is critical.

Purpose of the Study:

• To introduce a novel multi-stage, content-based framework for detecting profile cloning on LinkedIn.
• To classify profiles into four categories: legitimate, human-cloned, LLM-generated legitimate, and LLM-generated cloned.
• To address the inadequacy of traditional methods against AI-driven impersonation.

Main Methods:

• Semantic representation learning via attention-based section embedding aggregation.
• Linguistic style modeling using stylometric-perplexity features.
• Anomaly scoring with cluster-based outlier detection and ensemble classification through out-of-fold stacking.

Main Results:

• The proposed meta-ensemble model significantly outperforms baseline methods.
• Achieved macro-averaged accuracy, precision, recall, and F1-scores exceeding 96% on a 3,600-profile dataset.
• Demonstrated effectiveness in detecting both human-crafted and AI-generated impersonation.

Conclusions:

• A robust and scalable content-driven methodology for identity deception detection in OSNs is presented.
• Combining semantic, stylistic, and probabilistic signals is effective for detecting various forms of profile impersonation.
• The framework successfully addresses the challenge of AI-generated content in identity deception.