Target User Attack on Federated Recommender Systems with Proxy Items and Its Countermeasure

The decentralized nature of federated recommender systems (FedRecs) introduces new security risks, particularly target item promotion attacks, where attackers may attempt to manipulate the ranking and exposure rate of target items in users’ recommendation results. However, existing attacks typically promote each target item broadly to all users, which will compromise attack stealthiness while yielding only marginal gains in effectiveness. To address these issues, we propose a novel pro xy- it em based t arget u ser a ttack method on FedRecs, called ProitTUA , which selectively promotes each target item to a relevant user group rather than all users. The key novelty is to associate each target item with a relevant user group via category-wise proxy items. ProitTUA consists of two key stages: (1) item latent category inference, which infers latent item categories by leveraging auxiliary domain modeling and attack domain adaptation; and (2) proxy-item-based attack construction, which identifies proxy items for each category and then aligns each target item's embedding with its corresponding proxies. Extensive experiments show ProitTUA achieves superior attack effectiveness and stealthiness. To mitigate the potential threat posed by ProitTUA, we further design a defense mechanism based on gradient dynamics analysis. Experimental results demonstrate its effectiveness in reducing ProitTUA's impact on FedRecs.