Corrective unlearning in GNNs with Smarty for removing corrupted data
With the increasing vulnerability of Graph Neural Networks (GNNs) to data manipulation and adversarial attacks, there is a growing need for techniques capable of selectively removing the influence of corrupted information after training, while preserving model performance.
An approach to corrective unlearning in GNNs has been proposed in Smarty, enabling intervention even when only part of the manipulated data is known.
The method combines a phase of identifying influenced nodes—based on feature perturbation and analysis of prediction changes—with a correction phase in the embedding space. Representations are separated through opposite gradient updates, applying gradient ascent on nodes to be removed and gradient descent on clean nodes.
The results show the method’s ability to isolate corrupted information, restore performance in compromised regions of the graph, and maintain stability on unmanipulated data, without requiring full retraining.
Follow SMARTY for updates:
LinkedIn: https://www.linkedin.com/company/smartyeu/?viewAsMember=true
Website: https://www.smarty-project.eu