Methods and systems for detecting deepfakes
A system for detecting synthetic videos may include a server, a plurality of weak classifiers, and a strong classifier. The server may be configured to receive a prediction result from each of a plurality of weak classifiers; and send the prediction results from each of the plurality of weak classifiers to a strong classifier. The weak classifiers may be trained on real videos and known synthetic videos to analyze a distinct characteristic of a video file; detect irregularities of the distinct characteristic; generate a prediction result associated with the distinct characteristic, the prediction result being a prediction on whether the video file is synthetic; and output the prediction result to the server. The strong classifier may be trained to receive the prediction results of the plurality of weak classifiers from the server; analyze the prediction results; and determine if the video file is synthetic based on the prediction results.
1. A system for detecting fake videos comprising:
a server configured to:
receive a prediction result from each of a plurality of weak classifiers; and
send the prediction results from each of the plurality of weak classifiers to a strong classifier;
a plurality of weak classifiers, each weak classifier being trained on real videos and known fake videos to:
analyze a distinct characteristic of a video file;
detect irregularities of the distinct characteristic;
in response to detecting irregularities of the distinct characteristic, generate a prediction result dependent on the detected irregularities, wherein the prediction result is a prediction on whether the video file is fake; and
output the prediction result to the server; and
a strong classifier trained to:
receive the prediction result of each of the plurality of weak classifiers from the server;
analyze the prediction result from each of the plurality of weak classifiers; and
determine if the video file is fake based on the prediction results.
2. The system of claim 1 , wherein the prediction result is a numerical confidence level.
3. The system of claim 1 , wherein the known fake videos used to train the plurality of weak classifiers have had blurry frames removed.
4. The system of claim 1 , wherein a first weak classifier is trained to:
detect a mouth in the video file;
extract the mouth from the video file;
detect irregularities of the mouth, wherein irregularities may be associated with teeth or facial hair; and
generate a first prediction result based on the irregularities.
5. The system of claim 4 , wherein extracting the mouth from the video file comprises down-sampling the mouth.
6. The system of claim 4 , wherein extracting the mouth from the video file comprises:
extracting a pre-defined number of frames from the video file in which a mouth has been detected;
extracting a mouth from each extracted frame;
generating the first prediction result based on the mouth in each extracted frame; and
generating an average prediction result from the generated first prediction results.
7. The system of claim 1 , wherein a second weak classifier is trained to:
detect movement of a head in the video file;
calculate a pulse based on the detected movement;
detect irregularities of the pulse; and
generate a second prediction result associated with the irregularities of the pulse.
8. The system of claim 7 , wherein calculating the pulse based on the detected movement comprises:
identifying a plurality of features of the head;
decomposing trajectories of each feature into a set of component motions;
determining a component that best corresponds to a heartbeat based on its temporal frequency;
identifying peaks associated with the determined component; and
calculating the pulse based on the peaks.
9. The system of claim 1 , wherein a third weak classifier is trained to detect irregularities in audio gain of the video file.
10. A method for detecting fake videos comprising:
monitoring, by a server, a digital media source;
identifying, by the server, a video on the digital media source;
extracting, by the server, the video;
analyzing, by the server, the video with a plurality of weak classifiers, each weak classifier being trained to:
detect irregularities of a distinct characteristic of the video; and
in response to detecting irregularities of the distinct characteristic, generate a prediction result dependent on the detected irregularities, wherein the prediction result is a prediction on whether the video is fake;
analyzing, by the server, the prediction result of each of the plurality of weak classifiers with a strong classifier, the strong classifier being trained to determine if the video is fake based on the prediction results generated by each of the weak classifiers.
11. The method of claim 10 , wherein each weak classifier of the plurality of weak classifiers is trained on real videos and known fake videos.
12. The method of claim 11 , wherein the known fake videos used to train the plurality of weak classifiers have had blurry frames removed.
13. The method of claim 10 , wherein a first weak classifier is trained to:
detect a mouth in the video;
extract the mouth from the video;
detect irregularities of the mouth, wherein irregularities may be associated with teeth or facial hair; and
generate a first prediction result based on the irregularities.
14. The method of claim 13 , wherein extracting the mouth from the video comprises down-sampling the mouth.
15. The method of claim 13 , wherein extracting the mouth from the video comprises:
extracting a pre-defined number of frames from the video in which a mouth has been detected;
extracting a mouth from each extracted frame;
generating the first prediction result based on the mouth in each extracted frame; and
generating an average prediction result from the generated first prediction results.
16. The method of claim 10 , wherein a second weak classifier is trained to:
detect movement of a head in the video;
calculate a pulse based on the detected movement;
detect irregularities of the pulse; and
generate a second prediction result associated with the irregularities of the pulse.
17. The method of claim 16 , wherein calculating the pulse based on the detected movement comprises:
identifying a plurality of features of the head;
decomposing trajectories of each feature into a set of component motions;
determining a component that best corresponds to a heartbeat based on its temporal frequency;
identifying peaks associated with the determined component; and
calculating the pulse based on the peaks.
18. The method of claim 10 , wherein a third weak classifier is trained to detect irregularities in audio gain of the video.
19. The method of claim 10 , wherein the prediction result is a numerical confidence level.