Towards Backdoor Stealthiness in Model Parameter Space

Environment settings

check requirements.txt

Train a parameter space backdoor model

python train_backdoor.py --pr 0.1

pr is the poisoning rate of the target class.

Evaluate using CLP

python defense_lipschitzness_pruning.py

Generate UPGD

python generate_upgd.py --model_path ${clean_model_weights}