Exploring and Mitigating Adversarial Manipulation of Voting-Based Leaderboards

It is now common to evaluate Large Language Models (LLMs) by having humans manually vote to evaluate model outputs, in contrast to typical benchmarks that evaluate knowledge or skill at some particular task. Chatbot Arena, the most popular benchmark of this type, ranks models by asking users to select the better response between two randomly selected models (without revealing which model was responsible for the generations). These platforms are widely trusted as a fair and accurate measure of LLM capabilities. In this paper, we show that if bot protection and other defenses are not implemented, these voting-based benchmarks are potentially vulnerable to adversarial manipulation. Specifically, we show that an attacker can alter the leaderboard (to promote their favorite model or demote competitors) at the cost of roughly a thousand votes (verified in a simulated, offline version of Chatbot Arena). Our attack consists of two steps: first, we show how an attacker can determine which model was used to generate a given reply with more than $95\%$ accuracy; and then, the attacker can use this information to consistently vote for (or against) a target model. Working with the Chatbot Arena developers, we identify, propose, and implement mitigations to improve the robustness of Chatbot Arena against adversarial manipulation, which, based on our analysis, substantially increases the cost of such attacks. Some of these defenses were present before our collaboration, such as bot protection with Cloudflare, malicious user detection, and rate limiting. Others, including reCAPTCHA and login are being integrated to strengthen the security in Chatbot Arena.

The field of natural language processing has long relied on domain-specific, easy-to-implement evaluation metrics. But dramatic advances in LLM performance challenges traditional evaluation practices. As we show in this paper, moving from evaluations that use an objective source of truth to evaluations that utilize human inputs introduces the potential for new types of evaluation difficulties. We focus on this paper in validating one straightforward attack: by identifying and selectively voting for (or against) a particular model, an adversary can significantly alter the ordering of the best models.Mitigating this attack is feasible, and we are actively collaborating with the Chatbot Arena team to make Chatbot Arena more robust. We also encourage the community to explore and adopt mitigation strategies, such as voter authentication, rate limits, and more robust mechanisms for detecting malicious activities.