ReTrace

It includes training classifiers under different unlearning settings, extracting traces, and reconstructing forgotten data with reinforcement learning and GANs.

Files

• unlearning.py – Train classifiers:
  • Normal training
  • Exact unlearning
  • Approximate unlearning
• trace_heatmap.py – Generate heatmaps of unlearning traces for forgotten classes.
• trace_distribution.py – Analyze and output the distribution of unlearning traces.
• RL_GAN.py – Use reinforcement learning with a pretrained GAN to reconstruct forgotten data.

Usage

Train $f^+$ (original model):

python unlearning.py --task train_fplus --dataset cifar100

Train $f^-$ (unlearned model):

python unlearning.py --task train_fminus_exact --epochs 120 --forget_classes 0 --dataset cifar100

Train RL-GAN:

python RL_GAN.py

Note: Pre-train the GAN and set the checkpoint path in RL_GAN.py before running.