Cost-Minimized Label-Flipping Poisoning Attack to LLM Alignment

Large language models (LLMs) are increasingly deployed in real-world systems, making it critical to understand their vulnerabilities. While data poisoning attacks during RLHF/DPO alignment have been studied empirically, their theoretical foundations remain unclear. We investigate the minimum-cost poisoning attack required to steer an LLM's policy toward an attacker's target by flipping preference labels during RLHF/DPO, without altering the compared outputs. We formulate this as a convex optimization problem with linear constraints, deriving lower and upper bounds on the minimum attack cost. As a byproduct of this theoretical analysis, we show that any existing label-flipping attack can be post-processed via our proposed method to reduce the number of label flips required while preserving the intended poisoning effect. Empirical results demonstrate that this cost-minimization post-processing can significantly reduce poisoning costs over baselines, particularly when the reward model's feature dimension is small relative to the dataset size. These findings highlight fundamental vulnerabilities in RLHF/DPO pipelines and provide tools to evaluate their robustness against low-cost poisoning attacks.

Policy Iteration for Pareto-Optimal Policies in Stochastic Stackelberg Games

In general-sum stochastic games, a stationary Stackelberg equilibrium (SSE) does not always exist, in which the leader maximizes leader's return for all the initial states when the follower takes the best response against the leader's policy. Existing methods of determining the SSEs require strong assumptions to guarantee the convergence and the coincidence of the limit with the SSE. Moreover, our analysis suggests that the performance at the fixed points of these methods is not reasonable when they are not SSEs. Herein, we introduced the concept of Pareto-optimality as a reasonable alternative to SSEs. We derive the policy improvement theorem for stochastic games with the best-response follower and propose an iterative algorithm to determine the Pareto-optimal policies based on it. Monotone improvement and convergence of the proposed approach are proved, and its convergence to SSEs is proved in a special case.