StealthRank introduces a sophisticated technique for manipulating LLM-based recommendation systems while maintaining textual fluency to avoid detection.

• Employs an energy-based optimization framework with Langevin dynamics to generate natural-looking manipulations
• Successfully optimizes both ranking position and stealth simultaneously
• Demonstrates practical attacks against real-world LLM-based recommendation systems
• Reveals critical security vulnerabilities in current LLM ranking mechanisms

This research highlights the urgent need for robust defenses in LLM-powered information retrieval systems, as malicious actors could manipulate product recommendations while evading current detection methods.

StealthRank: LLM Ranking Manipulation via Stealthy Prompt Optimization