An unlearning method that updates model weights in the opposite direction of poisoned data gradients to remove their influence.