gan-police/frequency-forensics

Deepfake detection using wavelet-packets in PyTorch, European Conference on Machine Learning (ECML PKDD) 2022.

frequency-forensics's Issues

No Wavelet decomposition during the training

It appears to me that we are always training on the raw normalized images and calculate the wavelet representation only in the validation loop. Am I missing something?

frequency-forensics/src/freqdect/train_classifier.py

Lines 104 to 108 in 149db53

	batch_images = batch['image'].cuda(non_blocking=True)
	batch_labels = batch['label'].cuda(non_blocking=True)
	batch_images = (batch_images - 112.52875) / 68.63312
	out = model(batch_images)

frequency-forensics/src/freqdect/train_classifier.py

Lines 31 to 43 in 149db53

	batch_images = val_batch['image'].cuda(non_blocking=True)
	batch_labels = val_batch['label'].cuda(non_blocking=True)
	# batch_labels = torch.nn.functional.one_hot(batch_labels)
	batch_images = (batch_images - 112.52875) / 68.63312
	if packets:
	channel_list = []
	for channel in range(3):
	channel_list.append(
	compute_pytorch_packet_representation_2d_tensor(
	batch_images[:, :, :, channel],
	wavelet_str=wavelet, max_lev=max_lev))
	batch_images = torch.stack(channel_list, -1)
	out = model(batch_images)

Reduce memory requirements in data preperation step

Make,

frequency-forensics/src/freqdect/prepare_dataset.py

Line 449 in dad3cc1

mean = torch.mean(img_data.double(), axis).float()

more memory efficient.

Application build

Is it possible to build the application to detect the fake images using the provided code

Final output

Can you share the screenshot of final output where the images are labeled as fake or real

Could you share the pretrained classiffer please?

Your work is enlightening. Could you share the pre-trained model of the classifier, please? Thank you.

Packet size does not match the model

Hi there, thanks for sharing the code for such a great work!

I am trying to train the network using my own dataset and I want to use a similar setup as CNN-ln-db4 in your paper.

I prepared the dataset using the following commends (all paths are omitted):

python -m freqdect.prepare_dataset --log-packets --wavelet db4 --level 3

My dataset contains images all in size 128 x 128.

And I run the training script as following:

python -m freqdect.train_classifier \
--data-prefix log_packets_db4_reflect_3 \
--features packets \
--nclasses 8 \
--model cnn

However, I got the following errors at line 147 of models.py.

RuntimeError: mat1 and mat2 shapes cannot be multiplied (512x1176 and 24x8)

By checking the code, I think this is because the input packet size is bigger than what is expected, so after the convolutions the output does not go down to 1 x 1.

I am wondering did I make any mistakes in preparing the dataset? What should I do to have packets with the correct size?

Thank you.

Log directory not created automatically

When training models, the resulting model file is stored in ./log/, which may not exist:

frequency-forensics/src/freqdect/train_classifier.py

Lines 264 to 272 in dad3cc1

	model_file = (
	"./log/"
	+ args.data_prefix.split("/")[-1]
	+ "_"
	+ str(args.model)
	+ "_"
	+ str(args.seed)
	+ ".pt"
	)

Folder should be created on demand, if not existent.

Complete code?

Is this a complete code of deepfake image detection using wavelet packets

preprocessing dataset with features=packets

For some reason when I preprocess with features=packets the outputs have shape=(128,128,3), the same as the raw images, but with features=log-packets they have a shape=(64,16,16,3)... as a result I get a shape mismatch when trying to train a CNN on packets... shouldn't packets and log-packets result in input features of the same shape?