Graph neural networks (GNNs) are effective at modeling relational data, but they have a known weakness: label poisoning. According to a new arXiv preprint, a small number of corrupted training labels can be enough to disrupt a GNN, because message passing spreads the resulting errors across the graph.
The paper proposes a defense called contrastive adversarial training. The idea is to train the model in a way that anticipates adversarial label corruption, using contrastive learning to make the representations more resilient. The authors position this as a targeted response to the specific failure mode of label poisoning in GNNs.
Because the abstract is truncated, the full experimental details are not available in the source. The main claim is clear, though: contrastive adversarial training can help GNNs withstand label poisoning rather than simply detecting it after the fact.