Identity verification method and apparatus, computer device and storage medium

Patent metadata

Field	Value
Publication number	US-12019718-B2
Application number	US-202117359125-A
Country	US
Kind code	B2
Filing date	Jun 25, 2021
Priority date	Apr 24, 2019
Publication date	Jun 25, 2024
Grant date	Jun 25, 2024

Abstract

An identity authentication method is provided, including: acquiring a raw feature of a user; calling an identity authentication model to extract a primary attribute feature vector in the raw feature, the primary attribute feature vector being an unbiased feature representation for selectively decoupling m−1 domain discrepancy features in the raw feature, and m being an integer greater than 2; and performing unbiased identity authentication based on the primary attribute feature vector to obtain an identity authentication result.

First claim

What is claimed is:

1. An identity authentication method, executed by a computer device, the method comprising:
acquiring a raw feature of a user, the raw feature containing m−1 domain discrepancy features, and m being an integer greater than 2;
extracting a primary attribute feature vector in the raw feature by:
transforming, by a basic generator, the raw feature into a global attribute feature vector; and
performing feature extraction, by a primary generator, on the global attribute feature vector to obtain the primary attribute feature vector, the primary attribute feature vector being an unbiased feature representation by selectively decoupling the m−1 domain discrepancy features in the raw feature; and
performing unbiased identity authentication based on the primary attribute feature vector to obtain an identity authentication result by a step comprising performing the identity authentication, by a primary discriminator, on the primary attribute feature vector to obtain the identity authentication result or performing first discrimination, by the primary discriminator, on the primary attribute feature vector to provide a combined attribute feature vector.

2. The method according to claim 1 , wherein the extracting the primary attribute feature vector in the raw feature comprises:
performing feature extraction on the raw feature, by an identity authentication model having a first generative adversarial network, to obtain the primary attribute feature vector in the raw feature.

3. The method according to claim 2 , further comprising training the first generative adversarial network by a step comprising:
when a first domain discrepancy feature and a second domain discrepancy feature having a causal relationship exist in the raw feature, ignoring decoupling learning with the first domain discrepancy feature during adversarial learning for the second domain discrepancy feature.

4. The method according to claim 2 , wherein the first generative adversarial network comprises m generators G 1 to G m , respectively corresponding to discriminators D 11 to D mm ;
each generator G j corresponds to m discriminators D j1 to D jm ;
the generator G j is configured to learn a feature of a j th attribute, the attribute comprising an identity and m−1 domains;
a generator G 1 corresponding to the identity is the primary generator, and a discriminator D 11 corresponding to the generator G 1 is the primary discriminator, wherein i, j, j′∈{1, 2, . . . , m};
wherein the method further comprising training the first generative adversarial network by steps comprising:
fixing all generators G 1 to G m , and optimizing all discriminators D ij to make an output of the first generative adversarial network approximate to a tag y i corresponding to an i th attribute;
fixing all discriminators D ij and optimizing all generators G 1 to G m to make an output of the first generative adversarial network approximate to a tag (1-y i ) corresponding to the i th attribute; and
alternately performing the two fixing operations above until a training end condition for generators G i and the discriminators D ij is satisfied,
wherein when a j′ th attribute and the j th attribute have a causal relationship, back-propagation of an output loss of discriminator D jj′ is skipped, wherein i, j, j′∈{1, 2, . . . , m}.

5. The method according to claim 4 , wherein the discriminators D 11 to D mm are divided into two categories: for all i, j∈{1, 2, . . . , m}, i≠j;
each discriminator D ii is configured to learn a feature of the i th attribute, and each discriminator D ij is configured to eliminate the feature of the j th attribute; and
learning of each discriminator D ii includes standard supervised learning, and learning of each discriminator D ij includes adversarial learning.

6. The method according to claim 1 , wherein the extracting the primary attribute feature vector in the raw feature comprises:
performing feature extraction on the raw feature by an identity authentication model having a first generative adversarial network and a second generative adversarial network, the first generative adversarial network being trained by selectively decoupling the m−1 domain discrepancy features based on a causal relationship, and the second generative adversarial network being trained by performing additive adversarial training on at least one random combination of attribute feature vectors of different attributes extracted by the first generative adversarial network, the attributes comprising an identity and m−1 domain discrepancies.

7. The method according to claim 6 , wherein the second generative adversarial network comprises a primary additive spatial transformer network and a primary recognition network;
the performing unbiased identity authentication based on the primary attribute feature vector to obtain the identity authentication result comprises:
converting, by the primary additive spatial transformer network, the combined attribute feature vector outputted by the first generative adversarial network to obtain an additive feature vector; and
performing identity recognition, by the primary recognition network, on the additive feature vector to obtain the identity authentication result.

8. The method according to claim 6 , further comprising training the second generative adversarial network by steps comprising:
randomly combining the attribute feature vectors extracted by the first generative adversarial network from a training set to obtain the combined attribute feature vector; and
performing additive adversarial training on the combined attribute feature vector,
wherein at least one attribute combination corresponding to the combined attribute feature vector is an attribute combination that does not appear in the training set.

9. The method according to claim 6 , wherein the second generative adversarial network comprises m additive spatial transformer networks and m recognition networks having one-to-one correspondence to m attributes, j∈{1, 2, . . . , m};
the method further comprises training the second generative adversarial network by the steps comprising:
randomly combining the attribute feature vectors corresponding to different attributes generated by the first generative adversarial network to generate n r combined attribute feature vectors;
dividing the n r combined attribute feature vectors into a first vector set and a second vector set, an attribute combination of the combined attribute feature vectors in the first vector set being an attribute combination appearing in a training set, and an attribute combination of the combined attribute feature vectors in the second vector set being an attribute combination that does not appear in the training set;
using the first vector set and the second vector set to predict the additive spatial transformer networks and the recognition networks, a j th additive spatial transformer network being configured to convert a j th combined attribute feature vector into a j th additive feature vector, and a j th recognition network being configured to perform tag recognition corresponding to the j th attribute on a sum feature vector of m additive feature vectors;
for a first loss of the first vector set generated when predicting the additive spatial transformer networks and the recognition networks, back-propagating the first loss to the recognition network and the additive spatial transformer network corresponding to each attribute; and
for a second loss of the second vector set generated when predicting the additive spatial transformer networks and the recognition networks, back-propagating the second loss to the recognition networks and the additive spatial transform

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Liang Jian
• Cao Yuren
• Zhang Chenbin
• Bai Kun

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/047

  Probabilistic or stochastic networks · CPC title