Rethinking the impact of noisy labels in graph classification: A utility and privacy perspective

De Li, Xianxian Li, Zeming Gan, Qiyu Li, Bin Qu, Jinyan Wang

TL;DR

This paper reveals that noisy labels in graph classification tasks don't just degrade model accuracy—they also significantly increase the risk of privacy leaks through membership inference attacks. To combat this, the authors introduce RGLC, a robust framework that uses principal component analysis and dual-view label correction to filter noise and protect data privacy without sacrificing performance.

What Problem Does This Paper Address?

Graph Neural Networks (GNNs) are powerful tools for classifying complex data like molecular structures or social networks. However, they are notoriously sensitive to "noisy labels"—training data where the category assigned to a graph is incorrect.

While previous research has focused on how noise hurts a model's utility (its accuracy), this paper highlights a critical second dimension: privacy. The authors demonstrate that noisy labels cause models to overfit, creating a "memorization" effect that makes it much easier for an attacker to determine if a specific data point was part of the training set (a Membership Inference Attack, or MIA). Most existing solutions for noisy labels are designed for images or node-level tasks and fail to address the unique structural complexities and privacy risks inherent in graph-level classification.

Methods

The authors propose the Robust Graph Neural Network Approach with Noisy Labeled Graph Classification (RGLC), which consists of three main stages:

  1. Two-Stage Noise Selection:
    • Stage 1: It uses the "small loss" criterion (the idea that models learn clean patterns before noisy ones) to identify a set of high-confidence samples.
  2. Stage 2: It performs Singular Value Decomposition (SVD) on these high-confidence samples to find the "first principal component vector" for each class. Clean samples typically align closely with this vector, while noisy ones do not, allowing for precise filtering.
  3. Dual-View Label Correction: Instead of just discarding noisy samples, RGLC tries to fix them. It combines information from two "views": the embedding space (how well a sample aligns with the class principal components) and the output space (the model's prediction on perturbed versions of the graph).
  4. Supervised Graph Contrastive Learning: Finally, the model uses the corrected labels to pull similar graphs together and push different ones apart in the embedding space. This makes the model's internal representations more robust and harder for privacy attackers to exploit.

Datasets

The RGLC method was evaluated on eight widely used graph classification benchmarks, spanning bioinformatics and social networks:

  • Bioinformatics: MUTAG, NCI1, PTC, COX2, PROTEINS, and PROTEINS_F.
  • Social Networks: IMDB-B and IMDB-M.

The researchers simulated noisy environments by randomly flipping labels at rates ranging from 10% to 50%.

Key Insights and Ideas

  • The Overfitting-Privacy Link: The paper confirms that noisy labels exacerbate overfitting. As shown in Figure 1, the gap between training and testing performance widens as noise increases, which directly translates to higher success rates for privacy attacks.
Accuracy and loss performance under different noise rates
  • Embeddings are More Robust than Outputs: In large datasets, the model's final output layer often becomes highly confused by noise. However, the authors found that the internal "embeddings" (the mathematical representations of the graphs) remain surprisingly discriminative. This is why RGLC focuses on the embedding space for noise filtering.
  • Contrastive Learning as a Shield: By enforcing "alignment and uniformity" in the embedding space through supervised contrastive learning, the model reduces the semantic variability that attackers use to identify training members.

Key Results

RGLC consistently outperformed twelve other state-of-the-art methods across all eight datasets.

  • Utility: At a 30% noise rate, RGLC achieved performance gains between 0.8% and 7.8% over the previous best method (OMG).
  • Privacy: The framework successfully reduced the accuracy of membership inference attacks to below 60%, significantly mitigating the privacy risks introduced by noise.

As illustrated in Figure 2, the RGLC method helps close the "accuracy gap" that typically signals a high risk of privacy leakage.

Membership inference attack scenarios and accuracy gaps

Limitations and Future Work

The authors identify two main areas for improvement:

  1. Class-Dependent Noise: The current study focuses on "class-independent" noise (random flips). In the real world, noise is often "class-dependent" (e.g., two similar-looking molecules are frequently mislabeled as each other), which is harder to filter.
  2. White-Box Attack Gap: While RGLC is highly effective against black-box attacks, defending against "white-box" attacks—where the attacker has access to the model's internal parameters—remains a challenge at very high noise levels (e.g., 50%).

Future research will focus on enhancing the model's robustness in these more extreme and complex noise scenarios.

Summary updated

Abstract

Graph neural networks based on message-passing mechanisms have achieved advanced results in graph classification tasks. However, their generalization performance degrades when noisy labels are present in the training data. Most existing noisy labeling approaches focus on the visual domain or graph node classification tasks and analyze the impact of noisy labels only from a utility perspective. Unlike existing work, in this paper, we measure the effects of noise labels on graph classification from data privacy and model utility perspectives. We find that noise labels degrade the model's generalization performance and enhance the ability of membership inference attacks on graph data privacy. To this end, we propose the robust graph neural network approach with noisy labeled graph classification. Specifically, we first accurately filter the noisy samples by high-confidence samples and the first feature principal component vector of each class. Then, the robust principal component vectors and the model output under data augmentation are utilized to achieve noise label correction guided by dual spatial information. Finally, supervised graph contrastive learning is introduced to enhance the embedding quality of the model and protect the privacy of the training graph data. The utility and privacy of the proposed method are validated by comparing twelve different methods on eight real graph classification datasets. Compared with the state-of-the-art methods, the RGLC method achieves at most and at least 7.8% and 0.8% performance gain at 30% noisy labeling rate, respectively, and reduces the accuracy of privacy attacks to below 60%.
BibTeX
@article{li2025rethinking,
  author = {De Li and Xianxian Li and Zeming Gan and Qiyu Li and Bin Qu and Jinyan Wang},
  title = {{Rethinking the impact of noisy labels in graph classification: A utility and privacy perspective}},
  journal = {Neural Networks},
  year = {2025},
  doi = {10.1016/j.neunet.2024.106919},
}
Figure