Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection

August 15, 2024 · View on GitHub

Beijing Jiaotong University, YanShan University, A*Star

Reference github repository for the paper Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection.

@misc{tan2023rethinking,
      title={Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection}, 
      author={Chuangchuang Tan and Huan Liu and Yao Zhao and Shikui Wei and Guanghua Gu and Ping Liu and Yunchao Wei},
      year={2023},
      eprint={2312.10461},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

News 🆕

2024/02: NPR is accepted by CVPR 2024! Congratulations and thanks to my all co-authors!
2024/05: 🤗Online Demo

Environment setup

Classification environment: We recommend installing the required packages by running the command:

pip install -r requirements.txt

In order to ensure the reproducibility of the results, we provide the following suggestions：

Docker image: nvcr.io/nvidia/tensorflow:21.02-tf1-py3
Conda environment: ./pytorch18/bin/python
Random seed during testing period: Random seed

Getting the data

	paper	Url
Train set	CNNDetection CVPR2020	Baidudrive
Val set	CNNDetection CVPR2020	Baidudrive
Table1 Test	CNNDetection CVPR2020	Baidudrive
Table2 Test	FreqNet AAAI2024	googledrive
Table3 Test	DIRE ICCV2023	googledrive
Table4 Test	UniversalFakeDetect CVPR2023	googledrive
Table5 Test	Diffusion1kStep	googledrive

pip install gdown==4.7.1

chmod 777 ./download_dataset.sh

./download_dataset.sh

Directory structure

Click to expand the folder tree structure.

datasets
|-- ForenSynths_train_val
|   |-- train
|   |   |-- car
|   |   |-- cat
|   |   |-- chair
|   |   `-- horse
|   `-- val
|   |   |-- car
|   |   |-- cat
|   |   |-- chair
|   |   `-- horse
|   |-- test
|       |-- biggan
|       |-- cyclegan
|       |-- deepfake
|       |-- gaugan
|       |-- progan
|       |-- stargan
|       |-- stylegan
|       `-- stylegan2
`-- Generalization_Test
    |-- ForenSynths_test       # Table1
    |   |-- biggan
    |   |-- cyclegan
    |   |-- deepfake
    |   |-- gaugan
    |   |-- progan
    |   |-- stargan
    |   |-- stylegan
    |   `-- stylegan2
    |-- GANGen-Detection     # Table2
    |   |-- AttGAN
    |   |-- BEGAN
    |   |-- CramerGAN
    |   |-- InfoMaxGAN
    |   |-- MMDGAN
    |   |-- RelGAN
    |   |-- S3GAN
    |   |-- SNGAN
    |   `-- STGAN
    |-- DiffusionForensics  # Table3
    |   |-- adm
    |   |-- ddpm
    |   |-- iddpm
    |   |-- ldm
    |   |-- pndm
    |   |-- sdv1_new
    |   |-- sdv2
    |   `-- vqdiffusion
    `-- UniversalFakeDetect # Table4
    |   |-- dalle
    |   |-- glide_100_10
    |   |-- glide_100_27
    |   |-- glide_50_27
    |   |-- guided          # Also known as ADM.
    |   |-- ldm_100
    |   |-- ldm_200
    |   `-- ldm_200_cfg
    |-- Diffusion1kStep     # Table5
        |-- DALLE
        |-- ddpm
        |-- guided-diffusion    # Also known as ADM.
        |-- improved-diffusion  # Also known as IDDPM.
        `-- midjourney

Training the model

CUDA_VISIBLE_DEVICES=0 ./pytorch18/bin/python train.py --name 4class-resnet-car-cat-chair-horse --dataroot ./datasets/ForenSynths_train_val --classes car,cat,chair,horse --batch_size 32 --delr_freq 10 --lr 0.0002 --niter 50

Testing the detector

Modify the dataroot in test.py.

CUDA_VISIBLE_DEVICES=0 ./pytorch18/bin/python test.py --model_path ./NPR.pth  --batch_size {BS}

Detection Results

AIGCDetectBenchmark using ProGAN-4class checkpoint

When testing on AIGCDetectBenchmark, set no_resize and no_crop to True, and set batch_size to 1. To deal with images of odd sizes, add the following code in network/resnet.py.

n,c,w,h = x.shape
if w%2 == 1 : x = x[:,:,:-1,:]
if h%2 == 1 : x = x[:,:,:,:-1]

Generator	CNNSpot	FreDect	Fusing	GramNet	LNP	LGrad	DIRE-G	DIRE-D	UnivFD	RPTCon	NPR
ProGAN	100.00	99.36	100.00	99.99	99.67	99.83	95.19	52.75	99.81	100.00	99.9
StyleGan	90.17	78.02	85.20	87.05	91.75	91.08	83.03	51.31	84.93	92.77	96.1
BigGAN	71.17	81.97	77.40	67.33	77.75	85.62	70.12	49.70	95.08	95.80	87.3
CycleGAN	87.62	78.77	87.00	86.07	84.10	86.94	74.19	49.58	98.33	70.17	90.3
StarGAN	94.60	94.62	97.00	95.05	99.92	99.27	95.47	46.72	95.75	99.97	99.6
GauGAN	81.42	80.57	77.00	69.35	75.39	78.46	67.79	51.23	99.47	71.58	85.4
Stylegan2	86.91	66.19	83.30	87.28	94.64	85.32	75.31	51.72	74.96	89.55	98.1
WFIR	91.65	50.75	66.80	86.80	70.85	55.70	58.05	53.30	86.90	85.80	60.7
ADM	60.39	63.42	49.00	58.61	84.73	67.15	75.78	98.25	66.87	82.17	84.9
Glide	58.07	54.13	57.20	54.50	80.52	66.11	71.75	92.42	62.46	83.79	96.7
Midjourney	51.39	45.87	52.20	50.02	65.55	65.35	58.01	89.45	56.13	90.12	92.6
SDv1.4	50.57	38.79	51.00	51.70	85.55	63.02	49.74	91.24	63.66	95.38	97.4
SDv1.5	50.53	39.21	51.40	52.16	85.67	63.67	49.83	91.63	63.49	95.30	97.5
VQDM	56.46	77.80	55.10	52.86	74.46	72.99	53.68	91.90	85.31	88.91	90.1
Wukong	51.03	40.30	51.70	50.76	82.06	59.55	54.46	90.90	70.93	91.07	91.7
DALLE2	50.45	34.70	52.80	49.25	88.75	65.45	66.48	92.45	50.75	96.60	99.6
Average	70.78	64.03	68.38	68.67	83.84	75.34	68.68	71.53	78.43	89.31	91.7

GenImage

(1) Change "resize" to "translate and duplicate". (2) Set random seed to 70. (3) During testing, set no_crop to False.

(1)

dset = datasets.ImageFolder(
    root,
    transforms.Compose([
        # rz_func,
	transforms.Lambda(lambda img: translate_duplicate(img, opt.cropSize)),
	crop_func,
	flip_func,
	transforms.ToTensor(),
	transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]))

import math
def translate_duplicate(img, cropSize):
    if min(img.size) < cropSize:
        width, height = img.size
        
        new_width = width * math.ceil(cropSize/width)
        new_height = height * math.ceil(cropSize/height)
        
        new_img = Image.new('RGB', (new_width, new_height))
        for i in range(0, new_width, width):
            for j in range(0, new_height, height):
                new_img.paste(img, (i, j))
        return new_img
    else:
        return img

(2) Set random seed to 70.

(3) During testing, set no_crop to False. And set test config

vals =       ['ADM', 'biggan', 'glide', 'midjourney', 'sdv5', 'vqdm', 'wukong']
multiclass = [ 0,     0,        0,       0,            0,      0,      0      ]

./pytorch18/bin/python  train.py --dataroot {GenImage Path} --name sdv4_bs32_ --batch_size 32 --lr 0.0002 --niter 1  --cropSize 224 --classes sdv4

Train with sdv4 as the training set, using a random seed of 70. Pretrained checkpoint.

Generator	Acc.	A.P.
ADM	87.8	96.0
biggan	80.7	89.8
glide	93.2	99.1
midjourney	91.7	97.9
sdv5	94.4	99.9
vqdm	88.7	96.1
wukong	94.0	99.7
Mean	90.1	96.9

Acknowledgments

This repository borrows partially from the CNNDetection.