Abstract
Noisy labels in graph classification present a dual challenge: they degrade model accuracy and increase vulnerability to membership inference attacks (MIA). This paper introduces RGLC (Robust Graph neural network with noisy Labeled graph Classification), a framework designed to restore model utility and enhance privacy. By leveraging principal component analysis in the embedding space and a dual-view label correction mechanism, RGLC effectively filters noise and prevents the overfitting that leads to privacy leaks.
Introduction: The Hidden Cost of Noisy Labels
In domains like bioinformatics and social network analysis, graph data labels are frequently inaccurate due to human error or automated heuristics. While the impact of label noise on image classification is well-documented, its effect on graph-level classification has remained under-explored. This research identifies a critical "double-edged sword": noisy labels not only harm generalization but also force Graph Neural Networks (GNNs) to overfit on erroneous data. This overfitting creates a significant "privacy gap"—a measurable difference in model behavior between training and unseen data—which attackers exploit to identify specific training samples through Membership Inference Attacks (MIA).
The Privacy-Utility Gap
The core insight of this work is the direct link between label noise and privacy liability. As noise increases, the model works harder to "memorize" exceptions, leaving a digital fingerprint. This manifests as a widening gap between training and test accuracy, providing a clear signal for MIA attackers.
The RGLC Framework
To address these issues, the authors propose RGLC, which employs a three-step pipeline to handle noise robustly:
- Two-Stage Noise Filtering: RGLC first applies the "small loss criterion," identifying high-confidence samples that the model learns before memorizing noise. It then constructs a Gram matrix in the embedding space and uses Singular Value Decomposition (SVD) to isolate "clean" samples that align with the class's first principal component.
- Dual-View Label Correction: Instead of discarding noisy samples, RGLC attempts to correct them by reconciling information from the embedding space (proximity to principal components) and the output space (predictions from a noise-perturbed encoder).
- Supervised Graph Contrastive Learning (SCL): Using corrected labels, the model pulls similar graphs together in the feature space. This tightening of clusters reduces the semantic variability that MIA attackers rely on, simultaneously improving both robustness and privacy.
Key Results and Insights
Evaluated on eight benchmarks (including MUTAG and NCI1), RGLC demonstrated superior performance:
- Utility: At a 30% noise rate, RGLC outperformed the previous state-of-the-art (OMG) by up to 7.8% in accuracy.
- Privacy: The framework reduced MIA accuracy to below 60%, approaching the 50% baseline of a random guess, effectively closing the privacy gap.
- Embedding Reliability: A key finding is that internal embedding spaces are more resilient to noise than final softmax outputs, making them a more reliable foundation for noise filtering.
Conclusion and Future Directions
RGLC provides a robust defense against the utility degradation and privacy risks inherent in noisy graph data. Future work will focus on addressing class-dependent noise—where specific categories are systematically confused—and closing the remaining vulnerability gap in white-box attack scenarios where internal embeddings are fully exposed to attackers.
Summary updated