Ethics Privacy
A new method to verify if AI has truly forgotten prohibited data
A new method aims to make graph neural networks more transparent when they are required to forget data for privacy reasons. The background includes, for example, the EU's General Data Protection Regulation (GDPR), which gives individuals the right to request the deletion of their data - including from AI models trained with this data.
Graph neural networks process network-structured data, such as social media user networks or molecular structures. Removing data from them, known as unlearning, has proven difficult: current unlearning methods are primarily designed for efficiency and scale, but their effectiveness is hard to verify in black-box-like models.
The method introduced by Imran Ahsan, Hyunwook Yun, Jinsung Kim, and Mucheol Kim does not retrain the model itself but acts as a separate verifier. It captures model snapshots before and after the removal process and compares them using explainability methods. The idea is that if information related to a specific node or edge has truly been forgotten, its impact on the model's decisions and internal explanations will change visibly.
The verifier utilizes five explainability metrics, including so-called residual attribution, heatmap shift, and explainability score deviation. Additionally, it monitors local structural changes in the graph, measured, for example, as graph edit distance.
The goal of the method is to provide transparent, concrete evidence of whether forgetting has truly occurred, rather than relying solely on performance metrics. In this way, it can help both researchers and regulators assess whether graph neural networks comply with data deletion requirements.
Source: Forget and Explain: Transparent Verification of GNN Unlearning, ArXiv (AI).
Graph neural networks process network-structured data, such as social media user networks or molecular structures. Removing data from them, known as unlearning, has proven difficult: current unlearning methods are primarily designed for efficiency and scale, but their effectiveness is hard to verify in black-box-like models.
The method introduced by Imran Ahsan, Hyunwook Yun, Jinsung Kim, and Mucheol Kim does not retrain the model itself but acts as a separate verifier. It captures model snapshots before and after the removal process and compares them using explainability methods. The idea is that if information related to a specific node or edge has truly been forgotten, its impact on the model's decisions and internal explanations will change visibly.
The verifier utilizes five explainability metrics, including so-called residual attribution, heatmap shift, and explainability score deviation. Additionally, it monitors local structural changes in the graph, measured, for example, as graph edit distance.
The goal of the method is to provide transparent, concrete evidence of whether forgetting has truly occurred, rather than relying solely on performance metrics. In this way, it can help both researchers and regulators assess whether graph neural networks comply with data deletion requirements.
Source: Forget and Explain: Transparent Verification of GNN Unlearning, ArXiv (AI).
This text was generated with AI assistance and may contain errors. Please verify details from the original source.
Original research: Forget and Explain: Transparent Verification of GNN Unlearning
Publisher: ArXiv (AI)
Authors: Imran Ahsan, Hyunwook Yu, Jinsung Kim, Mucheol Kim
December 26, 2025
Read original →