Train/test split files

[plplo]

Thank you for sharing the code and dataset.

I would like to know when the Train/test split files can be made available.

thank you！

[hasamkhalid]

Hi, Thank you for your query. I believe the response below will answer your concerns related to the train/test split: Train/Test split Data Imbalance Issue: While splitting for train/test, we encountered a data imbalance issue. There are two ways of dealing with this data imbalance issue. I will explain both below: I. To run experiments, we performed frame-level detection. The number of frames extracted from Real videos was more than the number of frames extracted from Fake videos (Category B, C, D). For example, we have 500 Real videos and around 10,000 videos in Category C and D. So, we extracted ~140 frames from Real videos : 140 x 500 = 70,000 Frames. And we extracted ~7 frames from Fake videos : 7 x 10,000 = 70,000 Frames. (Note: This was done for each category of Fake video, i.e., B, C, D separately) By extracting frames this way, we got rid of the data imbalance issue. II. There is another way of coping with the data imbalance issue that we briefly mentioned in the papers as well. Since we collected Real videos from the VoxCeleb-2 dataset, the VoxCeleb dataset contains around a million Real videos. As we have fewer Real videos in our FakeAVCeleb dataset, to get rid of the data imbalance issue we can get more Real videos from the VoxCeleb dataset to train our models and run our experiments. Also, we have provided the actual ids of the persons that we selected from the VoxCeleb dataset and selected a single video per person, we can easily select more videos of the same person from the VoxCeleb dataset. III. Train/Test set split For the part where we said in the paper "test set contains 70 real and 70 fake videos", It means we selected 70 videos of 70 persons from a total of 500 persons. In the case of real, we had 70 videos, in the case of fake categories since we had more fake videos per person, we selected a single video per person. Resulting in 70 fake videos. Then we extracted frames as mentioned in my previous email. Moreover, in the case of audio, since each video has a single audio file, we converted the wave file into MFCC features, giving us a single MFCC per audio. Resulting in 70 real MFCC and 70 Fake MFCC. For more questions and clarity on dataset preprocessing and experiment settings-related questions, I would recommend you to read our paper mentioned below. This paper will give you more insights related to our baseline experiments. "Evaluation of an Audio-Video Multimodal Deepfake Dataset using Unimodal and Multimodal Detectors" https://dl.acm.org/doi/10.1145/3476099.3484315 I hope this email addresses your questions. If you want to know more you can write me anytime. Thank you. Regards, ------- Hasam Khalid, MS. College of Computing and Informatics Sungkyunkwan University, Suwon Campus, South Korea ***@***.***

…

On Thu, 7 Sep 2023, 10:37 am plplo, ***@***.***> wrote: Thank you for sharing the code and dataset. I would like to know when the Train/test split files can be made available. thank you！ — Reply to this email directly, view it on GitHub <#10>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANJHYYA7SPTQER5DLK6ORA3XZGIVDANCNFSM6AAAAAA4OTPB7Q> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[plplo]

Thank you for your detailed response, but may I ask if the train/test split files will be made public?

[Yizhichaoai]

Thank you for sharing the code and dataset.

I would like to know when the Train/test split files can be made available.

thank you！
Hello! Have you received the download link? I have submitted applications multiple times before, but I still haven't received the download link. If you have obtained it, could you kindly send me a mail?

[Yizhichaoai]

Hi, Thank you for your query. I believe the response below will answer your concerns related to the train/test split: Train/Test split Data Imbalance Issue: While splitting for train/test, we encountered a data imbalance issue. There are two ways of dealing with this data imbalance issue. I will explain both below: I. To run experiments, we performed frame-level detection. The number of frames extracted from Real videos was more than the number of frames extracted from Fake videos (Category B, C, D). For example, we have 500 Real videos and around 10,000 videos in Category C and D. So, we extracted ~140 frames from Real videos : 140 x 500 = 70,000 Frames. And we extracted ~7 frames from Fake videos : 7 x 10,000 = 70,000 Frames. (Note: This was done for each category of Fake video, i.e., B, C, D separately) By extracting frames this way, we got rid of the data imbalance issue. II. There is another way of coping with the data imbalance issue that we briefly mentioned in the papers as well. Since we collected Real videos from the VoxCeleb-2 dataset, the VoxCeleb dataset contains around a million Real videos. As we have fewer Real videos in our FakeAVCeleb dataset, to get rid of the data imbalance issue we can get more Real videos from the VoxCeleb dataset to train our models and run our experiments. Also, we have provided the actual ids of the persons that we selected from the VoxCeleb dataset and selected a single video per person, we can easily select more videos of the same person from the VoxCeleb dataset. III. Train/Test set split For the part where we said in the paper "test set contains 70 real and 70 fake videos", It means we selected 70 videos of 70 persons from a total of 500 persons. In the case of real, we had 70 videos, in the case of fake categories since we had more fake videos per person, we selected a single video per person. Resulting in 70 fake videos. Then we extracted frames as mentioned in my previous email. Moreover, in the case of audio, since each video has a single audio file, we converted the wave file into MFCC features, giving us a single MFCC per audio. Resulting in 70 real MFCC and 70 Fake MFCC. For more questions and clarity on dataset preprocessing and experiment settings-related questions, I would recommend you to read our paper mentioned below. This paper will give you more insights related to our baseline experiments. "Evaluation of an Audio-Video Multimodal Deepfake Dataset using Unimodal and Multimodal Detectors" [https://dl.acm.org/doi/10.1145/3476099.3484315 ](https://dl.acm.org/doi/10.1145/3476099.3484315) I hope this email addresses your questions. If you want to know more you can write me anytime. Thank you. Regards, ------- Hasam Khalid, MS. College of Computing and Informatics Sungkyunkwan University, Suwon Campus, South Korea @.***
…
On Thu, 7 Sep 2023, 10:37 am plplo, @.> wrote: Thank you for sharing the code and dataset. I would like to know when the Train/test split files can be made available. thank you！ — Reply to this email directly, view it on GitHub <#10>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANJHYYA7SPTQER5DLK6ORA3XZGIVDANCNFSM6AAAAAA4OTPB7Q . You are receiving this because you are subscribed to this thread.Message ID: @.>

Hello! Have you received the download link? I have submitted applications multiple times before, but I still haven't received the download link. If you have obtained it, could you kindly send me a mail?

[dlutskiv]

Hello @hasamkhalid !

I know it has been a while, but I wonder if there is a defined train/test split for FakeAVCeleb dataset? It would be a nice way to compare all various models on the same split! Thanks in advance.