GeniusHTX / Distillation

README

1. overview

This is the repo for our paper On the Effectiveness of Distillation in Mitigating Backdoors in Pre-trained Encoder.

The repo consists of 3 modules:

• DATASETS: the dataloader module, STL10, CIFAR10, GTSRB, SVHN are supported
• KDistill_ZOO: the distillation loss module, attention-based loss, layer-based loss and feature-based loss are supported.
• MODELS: the module to implement model architecture.
• EVALUATION: the module to evaluate the encoder performance.

2. Usage

python distillation.py --student <your_model_path> --teacher <your_model_path> --results_dir <your_saved_dir> --distill_way 'ATD'