When Graph Contrastive Learning Backfires: Spectral Vulnerability and Defense in Recommendation

Zongwei Wang, Min Gao, Junliang Yu, Shazia Sadiq, Hongzhi Yin, Ling Liu

Abstract

Graph Contrastive Learning (GCL) has demonstrated substantial promise in enhancing the robustness and generalization of recommender systems. However, we reveal an unexpected vulnerability: the integration of GCL inadvertently increases the susceptibility of a recommender to targeted promotion attacks. Through theoretical investigation and empirical validation, we identify the root cause as the spectral smoothing effect induced by contrastive optimization, which disperses item embeddings and unintentionally enhances the exposure of target items. We introduce CLeaR, a bi-level optimization attack method that amplifies spectral smoothness, and SIM, a spectral irregularity mitigation framework designed to detect and suppress targeted items. Extensive experiments demonstrate that GCL-based models exhibit greater susceptibility when evaluated with CLeaR, while SIM effectively mitigates these vulnerabilities.

Introduction

Recommender systems are increasingly targeted by promotion attacks, where adversaries inject fake user profiles to boost the visibility of specific items. While Graph Contrastive Learning (GCL) was initially thought to enhance model robustness through self-supervised tasks, this research identifies a critical "spectral" vulnerability. GCL models, despite their high accuracy, are significantly more susceptible to targeted attacks than traditional models like LightGCN. This paper explores why this "backfire" occurs and proposes a novel defense mechanism.

Methodology

Spectral Analysis and the Smoothing Effect

Using Singular Value Decomposition (SVD), the authors analyze the spectrum of item embeddings. Theoretical analysis of the standard GCL loss (InfoNCE) reveals that it encourages spectral smoothing. By spreading variance across more dimensions, GCL mitigates popularity bias but simultaneously brings unpopular ("cold") items closer to the center of the user distribution, making them easier to promote.

CLeaR: Exploiting Spectral Smoothness

To validate this vulnerability, the researchers developed CLeaR (Graph Contrastive Learning Recommendation Attack). This bi-level optimization framework uses an inner loop to train the recommender and an outer loop to craft malicious interactions that exploit spectral smoothness. This "white-box" approach serves as a rigorous stress test for GCL-based models.

SIM: Spectral Irregularity Mitigation

The proposed defense, SIM, operates in two stages:

  1. Detection: It identifies items that deviate from the smooth manifold of the embedding space by measuring reconstruction error when projected onto a low-rank subspace.
  2. Suppression: A mitigation loss penalizes the similarity between suspicious items and their most similar users, effectively removing them from recommendation lists without degrading overall performance.

Experimental Evaluation

The findings were evaluated on three benchmark datasets: DouBan (movie ratings), Epinions (consumer reviews), and Gowalla (location check-ins).

Key Results

The experiments confirmed that GCL-enhanced models (e.g., SGL, XSimGCL) consistently show higher HitRatio@50 (attack success) compared to LightGCN.

Recommendation performance comparison (Recall@50 and HitRatio@50) across DouBan, Epinions, and Gowalla datasets.

As shown in the figure above, while GCL models generally achieve higher Recall@50 (top row), they also exhibit significantly higher HitRatio@50 (bottom row) under attack, illustrating the vulnerability. The CLeaR attack proved more successful than existing baselines like GOAT or AUSH.

The SIM defense successfully reduced the attack's HitRatio by at least an order of magnitude on DouBan and Gowalla without compromising Recall@50. However, on the extremely sparse Epinions dataset, the defense was less effective, suggesting that distinguishing genuine rare interactions from malicious ones remains a challenge in sparse environments.

Discussion and Insights

The "double-edged sword" of GCL lies in its ability to disperse embeddings. While this prevents dimensional collapse and reduces popularity bias, it lowers the barrier for unpopular items to be recommended widely. Malicious items, however, leave a "spectral signature" due to their unnatural interaction patterns, which SIM exploits for detection.

Conclusion

This work highlights a critical security-performance trade-off in GCL-based recommendation. While GCL improves utility, its spectral smoothing effect creates a vulnerability to targeted promotion attacks. The proposed SIM framework offers a promising path forward by mitigating these risks through spectral analysis. Future work will focus on improving defense precision in sparse data environments and extending analysis to sequential recommendation models.

Summary updated

Abstract

Graph Contrastive Learning (GCL) has demonstrated substantial promise in enhancing the robustness and generalization of recommender systems, particularly by enabling models to leverage large-scale unlabeled data for improved representation learning. However, in this paper, we reveal an unexpected vulnerability: the integration of GCL inadvertently increases the susceptibility of a recommender to targeted promotion attacks. Through both theoretical investigation and empirical validation, we identify the root cause as the spectral smoothing effect induced by contrastive optimization, which disperses item embeddings across the representation space and unintentionally enhances the exposure of target items. Building on this insight, we introduce CLeaR, a bi-level optimization attack method that deliberately amplifies spectral smoothness, enabling a systematic investigation of the susceptibility of GCL-based recommendation models to targeted promotion attacks. Our findings highlight the urgent need for robust countermeasures; in response, we further propose SIM, a spectral irregularity mitigation framework designed to accurately detect and suppress targeted items without compromising model performance. Extensive experiments on multiple benchmark datasets demonstrate that, compared to existing targeted promotion attacks, GCL-based recommendation models exhibit greater susceptibility when evaluated with CLeaR, while SIM effectively mitigates these vulnerabilities.
BibTeX
@article{wang2025when,
  author = {Zongwei Wang and Min Gao and Junliang Yu and Shazia Sadiq and Hongzhi Yin and Ling Liu},
  title = {{When Graph Contrastive Learning Backfires: Spectral Vulnerability and Defense in Recommendation}},
  journal = {ACM Transactions on Information Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=RMglukH9sN},
}
Figure