Replay spoofing detection for automatic speaker verification system
Described herein are a system and techniques for detecting whether biometric data provided in an access request is genuine or a replay. In some embodiments, the system uses an machine learning model trained using genuine and replay sample data which is optimized in order to produce a result set in which results for the genuine samples are pulled closer to a genuine center and results for the replay samples are pushed away from the genuine center. Subjecting input biometric data (e.g., an audio sample) to the trained model results in a classification of the input biometric data as genuine or replay, which can then be used to determine whether or not to verify the input biometric data.
1. A method comprising:
training, by a service provider computer, a machine learning model by:
providing a set of genuine sample data;
providing a set of replay sample data; and
optimizing the machine learning model such that results of the set of genuine sample data are closer to a genuine center and the results of the set of replay sample data are further away from the genuine center;
receiving, by the service provider computer, input biometric data associated with a user;
subjecting, by the service provider computer, the input biometric data to the machine learning model resulting in a classification as genuine or replay; and
determining, by the service provider computer, whether or not to verify the input biometric data based on the classification,
wherein the machine learning model is a convolutional neural network,
wherein the convolutional neural network includes a one class loss function, and
wherein the one class loss function is a divisional one class center loss function, or a marginal one class center loss function,
wherein the divisional one class center loss function includes a loss DC :
ℒ
D
C
=
(
1
2
N
+
∑
y
i
=
0
x
i
-
c
2
2
)
(
1
2
N
-
∑
y
i
=
1
x
i
-
c
2
2
)
wherein the marginal one class center loss function includes a loss DC :
ℒ
D
C
=
1
2
∑
y
i
=
0
x
i
-
c
2
2
+
1
2
λ
∑
y
i
=
1
m
-
x
i
-
c
2
2
+
,
wherein (y i =0) indicates a genuine sample, c indicates a genuine center, (y i =1) indicates a replay sample, N + is a number of genuine samples, N − is a number of replay samples, x i is an embedding, m is a margin, and λ is a tradeoff parameter.
2. The method of claim 1 , wherein a layer including the one class loss function is connected to a fully connected layer.
3. The method of claim 1 , wherein the input biometric data is audio data which has not been transformed into a spectrum.
4. The method of claim 1 , wherein the classification is determined based on a distance between the genuine center and a result of the input biometric data.
5. The method of claim 4 , wherein determining whether or not to verify the input biometric data based on the classification comprises determining whether the result of the input biometric data is greater than some threshold distance from the genuine center.
6. The method of claim 1 , further comprising upon determining to verify the input biometric data, comparing the input biometric data to a biometric sample stored in relation to the user.
7. A service provider computer comprising
a processor;
a memory; and
a computer readable medium coupled to the processor, the computer readable medium comprising code, executable by the processor, to implement a method comprising:
training a machine learning model by:
providing a set of genuine sample data;
providing a set replay sample data; and
optimizing the machine learning model such that results of the set of genuine sample data are closer to a genuine center and the results of the set of replay sample data are further away from the genuine center,
receiving input biometric data associated with a user;
providing the input biometric data to the machine learning model as input resulting in a classification as genuine or replay; and
determining whether or not to verify the input biometric data based on the classification,
wherein the machine learning model is a convolutional neural network,
wherein the convolutional neural network includes a one class loss function, and
wherein the one class loss function is a divisional one class center loss function, or a marginal one class center loss function,
wherein the divisional one class center loss function includes a loss DC :
ℒ
D
C
=
(
1
2
N
+
∑
y
i
=
0
x
i
-
c
2
2
)
(
1
2
N
-
∑
y
i
=
1
x
i
-
c
2
2
)
wherein the marginal one class center loss function includes a loss DC :
ℒ
D
C
=
1
2
∑
y
i
=
0
x
i
-
c
2
2
+
1
2
λ
∑
y
i
=
1
m
-
x
i
-
c
2
2
+
,
and wherein (y i =0) indicates a genuine sample, c indicates a genuine center, (y i =1) indicates a replay sample, N + is a number of genuine samples, N − is a number of replay samples, x i is an embedding, m is a margin, and λ is a tradeoff parameter.
8. The service provider computer of claim 7 , wherein a layer including the one class loss function is connected to a fully connected layer.
9. The service provider computer of claim 7 , wherein the input biometric data is audio data which has not been transformed into a spectrum.
10. The service provider computer of claim 7 , wherein the classification is determined based on a distance between the genuine center and a result of the input biometric data.
11. The service provider computer of claim 10 , wherein determining whether or not to verify the input biometric data based on the classification comprises determining whether the result of the input biometric data is greater than some threshold distance from the genuine center.
12. The service provider computer of claim 7 , wherein the method further comprises upon determining to verify the input biometric data, comparing the input biometric data to a biometric sample stored in relation to the user.