LLM4MEA: Data-Free Model Extraction Attacks on Sequential Recommender Systems via Large Language Models

The researchers propose LLM4MEA, which uses a large language model as a human-like ranker to launch data-free model extraction attacks (MEA) on sequential recommender systems. The method generates training data through interactions between the LLM ranker and the target recommender system; the ranker analyzes historical interactions to understand user behavior and selects items with consistent preferences from the recommendation results to extend the interaction history. Experiments in the paper show that the method reduces the distribution discrepancy between synthetic data and real-world data by up to 64.98%, and improves model extraction attack performance by an average of 44.82%. The authors also propose a simple and effective defense strategy from a defensive perspective, and identify key hyperparameters of recommender systems that can mitigate the risk of model extraction attacks.