Data-driven audio deepfake detection
An exemplary method for generating a visual representation of manipulations in an audio signal includes inputting the audio signal into a trained machine-learning model, wherein the machine-learning model is trained by generating, based on a training bona fide audio signal, a training bona fide time-frequency representation; generating, based on a training spoofed audio signal, a training spoofed time-frequency representation, wherein the training spoofed audio signal is a manipulated version of the training bona fide audio signal; generating a training visual representation of manipulations in the training spoofed audio signal based at least on a difference between the training bona fide time-frequency representation and the training spoofed time-frequency representation; and training the audio deepfake detection machine-learning model based on the training visual representation of the manipulations in the training spoofed audio signal; and generating, by the machine-learning model, the visual representation of the manipulations in the audio signal.
1 . A method for generating a visual representation of manipulations in an audio signal, comprising:
inputting the audio signal into a trained machine-learning model, wherein the machine-learning model is trained by:
generating, based on a training bona fide audio signal, a training bona fide time-frequency representation;
generating, based on a training spoofed audio signal, a training spoofed time-frequency representation, wherein the training spoofed audio signal is a manipulated version of the training bona fide audio signal;
generating a training visual representation of manipulations in the training spoofed audio signal based at least on a difference between the training bona fide time-frequency representation and the training spoofed time-frequency representation; and
training the audio deepfake detection machine-learning model based on the training visual representation of the manipulations in the training spoofed audio signal; and
generating, by the machine-learning model, the visual representation of the manipulations in the audio signal.
2 . The method of claim 1 , further comprising displaying the visual representation of the manipulations in the audio signal.
3 . The method of claim 1 , further comprising displaying one or more timestamps associated with the manipulations in the audio signal.
4 . The method of claim 1 , further comprising:
inputting the visual representation of the manipulations in the audio signal into a language model; and
generating, by the language model, a natural language description of the manipulations in the audio signal.
5 . The method of claim 4 , wherein the language model is trained based on the training visual representation of the manipulations in the training spoofed audio signal and a natural-language description of the manipulations in the training spoofed audio signal.
6 . The method of claim 4 , further comprising displaying the natural language description of the manipulations in the audio signal.
7 . The method of claim 1 , wherein the audio signal comprises a real-time signal from a telephone call.
8 . The method of claim 7 , further comprising automatically terminating the telephone call.
9 . The method of claim 7 , further comprising displaying a warning that the telephone call comprises manipulations.
10 . The method of claim 7 , further comprising outputting a recommendation to terminate the telephone call.
11 . The method of claim 1 , wherein the audio signal comprises a recorded voicemail message.
12 . The method of claim 11 , further comprising displaying a warning that the recorded voicemail message comprises manipulations.
13 . The method of claim 1 , wherein the training bona fide time-frequency representation and the training spoofed time-frequency representation are spectrograms.
14 . The method of claim 1 , wherein the training bona fide audio signal comprises one or more utterances by a speaker and the training spoofed audio signal comprises a manipulated version of the one or more utterances by the speaker.
15 . The method of claim 1 , wherein the training spoofed audio signal is generated by a deepfake machine-learning model based on the training bona fide audio signal.
16 . The method of claim 1 , wherein the training bona fide audio signal and the training spoofed audio signal are time synchronized.
17 . The method of claim 1 , wherein generating the training visual representation of the manipulations in the training spoofed audio signal further comprises:
smoothing the training bona fide time-frequency representation; and
smoothing the training spoofed time-frequency representation, wherein the difference between the training bona fide time-frequency representation and the training spoofed time-frequency representation is calculated by subtracting the smoothed training bona fide time-frequency representation from the smoothed training spoofed time-frequency representation.
18 . The method of claim 1 , wherein generating the training visual representation of the manipulations in the training spoofed audio signal further comprises normalizing the difference between the training bona fide time-frequency representation and the training spoofed time-frequency representation.
19 . The method of claim 1 , wherein the machine-learning model comprises a diffusion model.
20 . The method of claim 1 , wherein the machine-learning model comprises a convolutional neural network.
21 . A system for generating a visual representation of manipulations in an audio signal, comprising: one or more processors; one or more memories; and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs including instructions for:
inputting the audio signal into a trained machine-learning model, wherein the machine-learning model is trained by:
generating, based on a training bona fide audio signal, a training bona fide time-frequency representation;
generating, based on a training spoofed audio signal, a training spoofed time-frequency representation, wherein the training spoofed audio signal is a manipulated version of the training bona fide audio signal;
generating a training visual representation of manipulations in the training spoofed audio signal based at least on a difference between the training bona fide time-frequency representation and the training spoofed time-frequency representation; and
training the audio deepfake detection machine-learning model based on the training visual representation of the manipulations in the training spoofed audio signal; and
generating, by the machine-learning model, the visual representation of the manipulations in the audio signal.
22 . A non-transitory computer-readable storage medium storing one or more programs for generating a visual representation of manipulations in an audio signal, the one or more programs comprising instructions, which when executed by one or more processors of an electronic device, cause the electronic device to perform:
inputting the audio signal into a trained machine-learning model, wherein the machine-learning model is trained by:
generating, based on a training bona fide audio signal, a training bona fide time-frequency representation;
generating, based on a training spoofed audio signal, a training spoofed time-frequency representation, wherein the training spoofed audio signal is a manipulated version of the training bona fide audio signal;
generating a training visual representation of manipulations in the training spoofed audio signal based at least on a difference between the training bona fide time-frequency representation and the training spoofed time-frequency representation; and
training the audio deepfake detection machine-learning model based on the training visual representation of the manipulations in the training spoofed audio signal; and
generating, by the machine-learning model, the visual representation of the manipulations in the audio signal.