WE HAVE BEEN ACCEPTED AT ICCV2025!cheeeeeeeeeeers! soon we will upload more detailed and polished codes and more ckpts!

CIARD: Enhancing Accuracy and Robustness of Student Models through Cyclic Iterative Distillation

Instructions for Reproducing Results

1. Environment Setup
   Ensure you are using Python 3.8. Install all required packages using:

   pip install -r requirements.txt
   

2. Download Teacher Models

   • Download the clean teacher model checkpoint and place it in:
     models/nat_teacher_checkpoint/
   • Download the robust teacher model and place it accordingly.
     The models we used can be found here.

3. Dataset

   • Store the dataset in the data/ folder.

4. Run the Model

   • To run CIARD, use:

   python CIARD.py
   

   • You can modify the configuration in CIARD.py to change the student architecture or dataset.

   • To run evaluation, use:

   python attack_eval.py
   

   • You can（should) modify the configuration in attack_eval.py to set the student path.