RankGuard: 62x faster decentralized ranking immune to poisoning attacks

Online learning to rank (OLTR) traditionally relies on a central server to collect user interactions, giving operators power to introduce biases that conflict with user interests. Decentralized learning offers a trustless alternative, but it introduces a new vulnerability: malicious nodes can send poisoned model updates to degrade ranking quality. In a new paper, researchers present RankGuard, a framework that solves this by having each node evaluate incoming models against its own private click history, corrected for position bias. Only models that better explain a user's past interactions than the current local model are aggregated, making it fundamentally difficult for attackers to craft harmful updates that pass the test. The authors provide the first formal convergence guarantee for a decentralized OLTR algorithm, proving that honest nodes can converge to a high-quality ranking model.

RankGuard was tested against four poisoning attacks—including a powerful adaptive attack—using three different click models and four standard benchmarks. It outperformed all baselines in most scenarios, achieving up to 62x higher efficiency than its closest competitors. The system's lightweight communication overhead and robust defense mechanism make it practical for real-world decentralized applications, such as peer-to-peer search or federated recommendation systems. By eliminating the need for a central authority, RankGuard not only resists manipulation but also enhances privacy, since raw user interactions never leave the local device. This work marks a significant step toward trustworthy, user-centric ranking in distributed environments.