Creating machine learning models for detecting the application of specific deepfake tools
There is provided a computer implemented method of training a detection machine learning model (ML) for identifying a specific deepfake tool of a plurality of deepfake tools used to create a deepfake video, comprising: feeding a plurality of sample authentic videos into the specific deepfake tool, obtaining a plurality of deepfake videos as an outcome of the specific deepfake tool, creating a training dataset comprising a plurality of records, wherein a record includes a deepfake video labelled with a ground truth indicating the specific deepfake tool used to create the deepfake video, and training the detection ML model on the training dataset for detecting that the specific deepfake tool was used to create an input deepfake video.
1 . A computer implemented method of training a detection machine learning model (ML) for identifying a specific deepfake tool of a plurality of deepfake tools used to create a deepfake video, comprising:
feeding a plurality of sample authentic videos into the specific deepfake tool;
obtaining a plurality of deepfake videos as an outcome of the specific deepfake tool;
creating a training dataset comprising a plurality of records, wherein a record includes a deepfake video labelled with a ground truth indicating the specific deepfake tool used to create the deepfake video; and
training the detection ML model on the training dataset for detecting that the specific deepfake tool was used to create an input deepfake video;
creating a second plurality of records for inclusion in the training dataset for balancing the training dataset, the second plurality of records including a second plurality of sample authentic videos corresponding to the class, labelled with a ground truth indicating lack of use of the specific deepfake tool;
identifying a class of the specific deepfake tool according to at least one adaptation performed by the specific deepfake tool on an input video; and
selecting the plurality of sample authentic videos corresponding to the class;
wherein the plurality of deepfake videos are modified by the specific deepfake tool based on the class.
2 . The computer implemented method of claim 1 , wherein the plurality of sample videos are fed into the plurality of deepfake tools to obtain the plurality of deepfake videos, wherein the record includes the deepfake video labeled with the ground truth of the specific deepfake tool of the plurality of deepfake tools which was used to create the deepfake video, wherein the trained detection ML model is for detecting which specific deepfake tool of the plurality of deepfake tool was used to create the input deepfake video.
3 . The computer implemented method of claim 1 , wherein the class is selected from: adapting a face, adapting what a person says, adapting a body of a person, replacing an object, adapting a background, and removing an object.
4 . The computer implemented method of claim 1 , wherein the plurality of sample videos corresponding to the class are automatically identified by a class detector ML model trained to detect the presence of the class in an input video, trained on a training dataset of a plurality of training videos labelled with ground truth labels indicating presence of a feature associated with the class.
5 . A system for training a detection machine learning model (ML) for identifying a specific deepfake tool of a plurality of deepfake tools used to create a deepfake video, comprising:
at least one processor executing a code for:
feeding a plurality of sample authentic videos into the specific deepfake tool;
obtaining a plurality of deepfake videos as an outcome of the specific deepfake tool;
creating a training dataset comprising a plurality of records, wherein a record includes a deepfake video labelled with a ground truth indicating the specific deepfake tool used to create the deepfake video;
training the detection ML model on the training dataset for detecting that the specific deepfake tool was used to create an input deepfake video;
creating a second plurality of records for inclusion in the training dataset for balancing the training dataset, the second plurality of records including a second plurality of sample authentic videos corresponding to the class, labelled with a ground truth indicating lack of use of the specific deepfake tool;
identifying a class of the specific deepfake tool according to at least one adaptation performed by the specific deepfake tool on an input video; and
selecting the plurality of sample authentic videos corresponding to the class,
wherein the plurality of deepfake videos are modified by the specific deepfake tool based on the class.
6 . The system of claim 5 , wherein the class is selected from: adapting a face, adapting what a person says, adapting a body of a person, replacing an object, adapting a background, and removing an object.
7 . The system of claim 5 , wherein the plurality of sample videos corresponding to the class are automatically identified by a class detector ML model trained to detect the presence of the class in an input video, trained on a training dataset of a plurality of training videos labelled with ground truth labels indicating presence of a feature associated with the class.
8 . The system of claim 5 , wherein the plurality of sample videos are fed into the plurality of deepfake tools to obtain the plurality of deepfake videos, wherein the record includes the deepfake video labeled with the ground truth of the specific deepfake tool of the plurality of deepfake tools which was used to create the deepfake video, wherein the trained detection ML model is for detecting which specific deepfake tool of the plurality of deepfake tool was used to create the input deepfake video.
9 . A non-transitory medium storing program instructions for training a detection ML model for identifying a specific deepfake tool of a plurality of deepfake tools used to create a deepfake video, which when executed by at least one processor, cause the at least one processor to:
feed a plurality of sample authentic videos into the specific deepfake tool;
obtain a plurality of deepfake videos as an outcome of the specific deepfake tool;
create a training dataset comprising a plurality of records, wherein a record includes a deepfake video labelled with a ground truth indicating the specific deepfake tool used to create the deepfake video;
train the detection ML model on the training dataset for detecting that the specific deepfake tool was used to create an input deepfake video create a second plurality of records for inclusion in the training dataset for balancing the training dataset, the second plurality of records including a second plurality of sample authentic videos corresponding to the class, labelled with a ground truth indicating lack of use of the specific deepfake tool;
identify a class of the specific deepfake tool according to at least one adaptation performed by the specific deepfake tool on an input video; and
select the plurality of sample authentic videos corresponding to the class,
wherein the plurality of deepfake videos are modified by the specific deepfake tool based on the class.
10 . The non-transitory medium of claim 9 , wherein the class is selected from: adapting a face, adapting what a person says, adapting a body of a person, replacing an object, adapting a background, and removing an object.
11 . The non-transitory medium of claim 9 , wherein the plurality of sample videos corresponding to the class are automatically identified by a class detector ML model trained to detect the presence of the class in an input video, trained on a training dataset of a plurality of training videos labelled with ground truth labels indicating presence of a feature associated with the class.
12 . The non-transitory medium of claim 9 , wherein the plurality of sample videos are fed into the plurality of deepfake tools to obtain the plurality of deepfake videos, wherein the record includes the deepfake video labeled with the ground truth of the specific deepfake tool of the plurality of deepfake tools which was used to create the deepfake video, wherein the trained detection ML model is for detecting which specific deepfake tool of the plurality of deepfake tool was used to create the input deepfake video.