Vector Institute Researchers Develop Efficient Machine Unlearning Technique
Researchers at the Vector Institute for Artificial Intelligence, led by Nicolas Papernot, have developed a novel approach to machine unlearning, addressing the technical challenges of complying with digital privacy regulations like the EU's Right to be Forgotten. Traditional methods require retraining AI models from scratch when data is deleted, which is resource-intensive and slow. Papernot's team proposes a two-pronged strategy involving sharding and slicing data. By creating multiple smaller models that vote on predictions and establishing checkpoints during incremental data processing, specific user data can be removed by reverting to prior states without full retraining. This method significantly reduces the time and computational cost associated with data deletion requests. The research, titled Machine Unlearning, was accepted to the IEEE Symposium on Security and Privacy. It aims to provide a practical, algorithm-agnostic solution for organizations to handle privacy requests efficiently, ensuring that AI systems can adapt to legal requirements without compromising performance or incurring prohibitive operational delays.
Wire timeline
Vector Institute Researchers Develop Efficient Machine Unlearning Technique
Researchers at the Vector Institute for Artificial Intelligence, led by Nicolas Papernot, have developed a novel approach to machine unlearning, addressing the technical challenges of complying with digital privacy regulations like the EU's Right to be Forgotten. Traditional methods require retraining AI models from scratch when data is deleted, which is resource-intensive and slow. Papernot's team proposes a two-pronged strategy involving sharding and slicing data. By creating multiple smaller models that vote on predictions and establishing checkpoints during incremental data processing, specific user data can be removed by reverting to prior states without full retraining. This method significantly reduces the time and computational cost associated with data deletion requests. The research, titled Machine Unlearning, was accepted to the IEEE Symposium on Security and Privacy. It aims to provide a practical, algorithm-agnostic solution for organizations to handle privacy requests efficiently, ensuring that AI systems can adapt to legal requirements without compromising performance or incurring prohibitive operational delays.
Vector Institute for Artificial Intelligence