Reproduce “Exploring and Mitigating Adversarial Manipulation of Voting-Based Leaderboards” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yangsibo Huang, Milad Nasr, Anastasios Angelopoulos, Nicholas Carlini, Wei-Lin Chiang, Christopher A. Choquette-Choo, Daphne Ippolito, Matthew Jagielski and 5 more
Researchers from Google, UC Berkeley, and other institutions developed and tested adversarial manipulation techniques on voting-based LLM leaderboards, demonstrating that models can be de-anonymized with over 95% accuracy and rankings altered with under 5,000 votes. They also proposed and evaluated effective mitigation strategies.
Key insights
- LLMs exhibit distinct stylistic 'fingerprints' in their outputs, allowing for high-accuracy de-anonymization (over 95%) with basic text features at a low one-time cost.
- Voting-based leaderboards are highly vulnerable, with significant rank shifts (e.g., four positions) achievable with a relatively small number of adversarial votes (fewer than 5,000).
- Effective mitigations often involve a security-utility trade-off, particularly when perturbing publicly released rankings to deter sophisticated adversaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.