FLIP-KD

This repository contains the source code for the knowledge distillation of a fine-tuned CLIP model (FLIP) for the FAS (Face Anti-Spoofing) task.

Dataset Preparation

FLIP MCIO Dataset

We use the FLIP MCIO dataset for our experiments. Ensure the dataset is structured correctly and placed in the specified directories.

data/MCIO/frame/
|-- casia
    |-- train
    |   |--real
    |   |  |--1_1_frame0.png, 1_1_frame1.png 
    |   |--fake
    |      |--1_3_frame0.png, 1_3_frame1.png 
    |-- test
        |--real
        |  |--1_1_frame0.png, 1_1_frame1.png 
        |--fake
           |--1_3_frame0.png, 1_3_frame1.png 
|-- msu
    |-- train
    |   |--real
    |   |  |--real_client002_android_SD_scene01_frame0.png, real_client002_android_SD_scene01_frame1.png
    |   |--fake
    |      |--attack_client002_android_SD_ipad_video_scene01_frame0.png, attack_client002_android_SD_ipad_video_scene01_frame1.png
    |-- test
        |--real
        |  |--real_client001_android_SD_scene01_frame0.png, real_client001_android_SD_scene01_frame1.png
        |--fake
           |--attack_client001_android_SD_ipad_video_scene01_frame0.png, attack_client001_android_SD_ipad_video_scene01_frame1.png
|-- replay
    |-- train
    |   |--real
    |   |  |--real_client001_session01_webcam_authenticate_adverse_1_frame0.png, real_client001_session01_webcam_authenticate_adverse_1_frame1.png
    |   |--fake
    |      |--fixed_attack_highdef_client001_session01_highdef_photo_adverse_frame0.png, fixed_attack_highdef_client001_session01_highdef_photo_adverse_frame1.png
    |-- test
        |--real
        |  |--real_client009_session01_webcam_authenticate_adverse_1_frame0.png, real_client009_session01_webcam_authenticate_adverse_1_frame1.png
        |--fake
           |--fixed_attack_highdef_client009_session01_highdef_photo_adverse_frame0.png, fixed_attack_highdef_client009_session01_highdef_photo_adverse_frame1.png
|-- oulu
    |-- train
    |   |--real
    |   |  |--1_1_01_1_frame0.png, 1_1_01_1_frame1.png
    |   |--fake
    |      |--1_1_01_2_frame0.png, 1_1_01_2_frame1.png
    |-- test
        |--real
        |  |--1_1_36_1_frame0.png, 1_1_36_1_frame1.png
        |--fake
           |--1_1_36_2_frame0.png, 1_1_36_2_frame1.png
|-- celeb
    |-- real
    |   |--167_live_096546.jpg
    |-- fake
        |--197_spoof_420156.jpg       

Inference with FLIP Teacher Model

To perform inference with the FLIP teacher model, use the following command:

python infer.py \
    --report_logger_path path/to/save/performance.csv \
    --config M \
    --method flip_mcl \
    --ckpt /home/jiwon/FLIP_yy/student/t_ckpt/oulu_flip_mcl.pth.tar

Knowledge Distillation (Train Student Model)

To perform knowledge distillation, run the experiment.sh script with the following parameters:

sh experiment.sh

You can modify the parameter values in experiment.sh

--t_model_checkpoint "/home/jiwon/FLIP-KD/student/t_ckpt/msu_flip_mcl.pth.tar" \
--config custom \
--op_dir ckpt \
--report_logger_path logs \
--root "/home/jiwon/Dataset/FLIP_Dataset/MCIO/frame/" \
--dataroot "/home/jiwon/Dataset/FLIP_Dataset/MCIO/txt/" \
--epochs 100 \
--batch_size 3 \
--t_batch_size 30 \
--lr=0.000001 \
--alpha_ckd_loss 0. \
--alpha_fd_loss 0. \
--alpha_affinity_loss 0. \
--alpha_icl_loss 0.

Configuration Details

• Teacher Model: ViT-B-16
• Student Model: ViT-T-16
• Dataset: FLIP MCIO
• Preprocessing: Follow the same preprocessing steps as used for FLIP.