RSA-CR: Resisting Shilling Attacks in Citation Recommendation via Dumbbell Inductive Learning

Xiyue Gao, Yukai Liu, Zhuoqi Ma, Xiaotian Qiao, Hui Li, Chenyang Xu, KunHua Zhang, Jiangtao Cui

TL;DR

RSA-CR introduces a robust "Dumbbell Inductive Learning" framework to protect citation recommendation systems from "shilling attacks"—malicious attempts to game the system by injecting fake citations. By integrating author collaboration history, citation significance, and semantic content constraints, the model improves recommendation accuracy and robustness against attacks significantly compared to state-of-the-art baselines.

Introduction

In the academic world, citations are the currency of prestige. However, citation recommendation systems—tools that help researchers find relevant papers—are vulnerable to shilling attacks. This occurs when individuals or "citation rings" inject fake information, such as citing their own unrelated papers, to artificially inflate a paper's popularity and trick algorithms into recommending it.

The authors identify a specific challenge called "class solidification": the tendency for already-popular papers to stay popular due to researcher herd mentality. This makes it easier for attackers to "nuke" or "push" certain papers. Existing defense mechanisms for e-commerce often fail here because academic data is a complex mix of citation graphs, author networks, and deep textual content.

The RSA-CR Framework

The authors propose RSA-CR (Resisting Shilling Attacks in Citation Recommendation), which utilizes a novel Dumbbell Inductive Learning (DIL) architecture. The "dumbbell" name comes from the two-layer graph structure: one side focuses on authors and the other on papers, connected by authorship links.

The RSA-CR architecture featuring the Dumbbell Inductive Learning (DIL) framework.

The core of the defense lies in three strategies:

  1. Historical Collaborations: It measures the familiarity between authors. Shilling attackers often lack a genuine collaborative history with the authors of the papers they are targeting.
  2. Significant Citations: Not all citations are equal. The model weights citations based on the reputation of the citing author and the prestige of the journal or conference, making it harder for low-quality "shill" papers to influence the results.
  3. Content Constraints: Using doc2vec, the model ensures that a recommended paper is semantically related to the manuscript. This prevents attackers from successfully linking unrelated papers across different domains.

The learning process is inductive, meaning the model can generate embeddings for newly published papers (unseen during training) by aggregating features from their neighbors, which is essential for the fast-paced world of academic publishing.

Experimental Evaluation

The model was evaluated on four major academic datasets: ACL-OCL, unarXive, AMiner, and Saaca (Shilling Attack Academic Dataset). Saaca is a newly introduced dataset that uses retracted papers (withdrawn due to fraudulent peer review or citation manipulation) as a real-world proxy for shilling attacks.

Performance comparison of RSA-CR against baseline models across four academic datasets.

Key Results

  • Superior Accuracy: RSA-CR consistently outperformed six baselines across all datasets. As shown in the performance table, it achieved an HR@5 (Hit Ratio) nearly double that of traditional graph-based methods like LINE in several instances.
  • Robustness: On the Saaca dataset, RSA-CR achieved an Attack Hit Ratio (AHR) of 0.0044 at top-1 recommendations, significantly lower than Paper2vec (0.3668). This indicates that even when the system is under attack, RSA-CR is highly resistant to recommending "shill" papers.
  • Ablation Success: Ablation studies confirmed that the "Significant Citations" component is the most critical factor in resisting attacks, as its removal caused the largest drop in performance.

Conclusion

By performing aggregation on both the author side and the paper side, RSA-CR "anchors" a paper's credibility. An author's historical reputation directly influences the "strength" of the paper's aggregation, making it much harder to spoof relevance through citations alone. While the dual-layer aggregation is computationally intensive, it provides a necessary layer of security for the integrity of academic recommendations.

Summary updated

Abstract

Citation recommendation aims to provide researchers with the most relevant references for their manuscripts, helping them swiftly discover pertinent studies and bolster the reliability of their arguments. However, some individuals manipulate these recommendation systems by injecting false information, such as deliberately inflating the citation count of their own papers, to obtain favorable recommendations and ratings. This form of attack, commonly termed “shilling attack”, is not only highly concealed but also has an unimaginable impact on all scientific research. To address this problem, we theoretically reveal the impact of shilling attacks on citation recommendation and propose three feasible resistance strategies: historical collaborations, significant citations and content constraints. Based on these insights, we introduce RSA-CR, a robust and hybrid citation recommendation algorithm resistant to shilling attacks. The algorithm constructs a two-layer academic graph and uses random and content generation strategies to initialize author and paper embeddings. Confidence-guided inductive aggregations based on collaboration and citation relationships are then performed at the author and paper sides, where author aggregation results directly influences the paper aggregation strength. Finally, recommendations are made by measuring the distances between the fused paper embeddings. The entire learning process resembles a dumbbell, hence termed “dumbbell inductive learning”. Experiments on four academic datasets demonstrate that our method outperforms baselines in both effectiveness and robustness.
BibTeX
@inproceedings{gao2026rsa,
  author = {Xiyue Gao and Yukai Liu and Zhuoqi Ma and Xiaotian Qiao and Hui Li and Chenyang Xu and KunHua ZHANG and Jiangtao Cui},
  title = {{RSA-CR: Resisting Shilling Attacks in Citation Recommendation via Dumbbell Inductive Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i42.40854},
}
Figure