Reference Model

How it works

Benefits

Highlighted claims

• The reference model serves as the center of an L2-norm projection ball with radius delta. — Asynchronous Federated Unlearning with Invariance Calibration for Medical Imaging
• The reference model is computed by subtracting the target client's cached previous contribution from the current global model. — Asynchronous Federated Unlearning with Invariance Calibration for Medical Imaging
• The reference model anchors the unlearning trajectory to reduce drift and avoid catastrophic forgetting of retained knowledge. — Asynchronous Federated Unlearning with Invariance Calibration for Medical Imaging
• The projection ball is intended to keep destructive gradient ascent near a plausible retained model. — Asynchronous Federated Unlearning with Invariance Calibration for Medical Imaging