DEEPFAKE DETECTION
Disclosed are systems and methods including software processes executed by a server that detect audio-based synthetic speech (“deepfakes”) in a call conversation. The server applies an NLP engine to transcribe call audio and analyze the text for anomalous patterns to detect synthetic speech. Additionally or alternatively, the server executes a voice “liveness” detection system for detecting machine speech, such as synthetic speech or replayed speech. The system performs phrase repetition detection, background change detection, and passive voice liveness detection in call audio signals to detect liveness of a speech utterance. An automated model update module allows the liveness detection model to adapt to new types of presentation attacks, based on the human provided feedback.
1 . A computer-implemented method for detecting machine-based speech in calls, comprising:
obtaining, by a computer, a raw audio signal from a calling device including a speech signal for a speaker;
applying, by the computer, a spoofprint extractor of a machine-learning architecture on the raw audio signal to extract spoofprint embedding representing a set of spoofing artifacts in a set of acoustic features;
applying, by the computer, a spoofing classifier of the machine-learning architecture on the spoofprint embedding to generate a liveness score indicating a likelihood that the speaker is a human speaker; and
retraining, by the computer, at least a portion of the machine-learning architecture in response to identifying a new machine spoof attack.
2 . The method of claim 1 , further comprising receiving, by the computer, a user input containing an indication, via a user interface, indicating the speech signal in the raw audio signal as the new machine spoof attack.
3 . The method of claim 1 , further comprising determining, by the computer, a loss metric to retrain the machine-learning architecture based on a comparison between the liveness score and an indication of the new machine spoof attack.
4 . The method of claim 1 , further comprising updating, by the computer, the machine-learning architecture in response to identifying no machine spoof attack associated with the speech signal in the raw audio signal.
5 . The method of claim 1 , further comprising generating, by the computer, the set of spoofing artifacts in the set of acoustic features from the raw audio signal.
6 . The method of claim 1 , further comprising training, by the computer, the spoofprint extractor and the spoofing classifier of the machine-learning architecture using an initial training dataset comprising a plurality of examples, each of the plurality of examples identifying (i) a sample raw audio signal and (ii) an indication of one of human speaker or machine spoof attack.
7 . The method of claim 1 , wherein retraining the machine-learning architecture further comprises retraining both of the spoofprint extractor and the spoofing classifier of the machine-learning architecture using an identification of the new machine spoof attack.
8 . The method of claim 1 , wherein retraining the machine-learning architecture further comprises retraining the portion of the machine-learning architecture using an identification of the new machine spoof attack, while maintaining a remaining portion of the machine-learning architecture.
9 . The method of claim 1 , further comprising updating, by the computer, a plurality of data points of a training dataset to include a new data point associated with the identification of the new machine spoof attack, and
wherein computer retrains the machine-learning architecture in accordance with one or more policies and using the plurality of data points with the new data point.
10 . The method of claim 1 , further comprising generating, by the computer, an indicator, for a user interface, indicating the speaker is one of a human speaker or a machine spoof attack, based upon a comparison between the liveness score and a threshold.
11 . A system for detecting machine speech in calls, comprising:
a computer comprising one or more processors configured to:
obtain a raw audio signal from a calling device including a speech signal for a speaker;
apply a spoofprint extractor of a machine-learning architecture on the raw audio signal to extract spoofprint embedding representing a set of spoofing artifacts in a set of acoustic features;
apply a spoofing classifier of the machine-learning architecture on the spoofprint embedding to generate a liveness score indicating a likelihood that the speaker is a human speaker; and
retrain at least a portion of the machine-learning architecture in response to identifying a new machine spoof attack.
12 . The system of claim 11 , wherein the computer is further configured to receive a user input containing an indicator, via a user interface, indicating the speech signal in the raw audio signal as the new machine spoof attack.
13 . The system of claim 11 , wherein the computer is further configured to determine a loss metric to retrain the machine-learning architecture based on a comparison between the liveness score and an indication of the new machine spoof attack.
14 . The system of claim 11 , wherein the computer is further configured to update the machine-learning architecture in response to identifying no machine spoof attack associated with the speech signal in the raw audio signal.
15 . The system of claim 11 , wherein the computer is further configured to generate the set of spoofing artifacts in the set of acoustic features from the raw audio signal.
16 . The system of claim 11 , wherein the computer is further configured to train the spoofprint extractor and the spoofing classifier of the machine-learning architecture using an initial training dataset comprising a plurality of examples, each of the plurality of examples includes (i) a sample raw audio signal and (ii) an indicator that the examples contains one of a human speaker or a machine spoof attack.
17 . The system of claim 11 , wherein, when retraining the machine-learning architecture, the computer is further configured to retrain the spoofprint extractor and the spoofing classifier of the machine-learning architecture using an identification of the new machine spoof attack.
18 . The system of claim 11 , wherein, when retraining the machine-learning architecture, the computer is further configured to retrain the portion of the machine-learning architecture using an indication of the new machine spoof attack, and wherein a remaining portion of the machine-learning architecture is fixed.
19 . The system of claim 11 , wherein the computer is further configured to update to a plurality of data points of a training dataset to include a new data point associated with the identification of the new machine spoof attack; and
wherein, when retraining the machine-learning architecture, the computer is further configured to retrain the machine-learning architecture in accordance with one or more policies, using the plurality of data points with the new data point.
20 . The system of claim 11 , the computer is further configured to generate an indicator, for a user interface, indicating the speaker as one of a human speaker or a machine spoof attack, based upon a comparison between the liveness score and a threshold.