Title:
FEATURE MANIPULATION-BASED ATTACK AND DEFENSE METHOD FOR CONTINUOUS LEARNING ABILITY SYSTEM
Document Type and Number:
WIPO Patent Application WO/2023/070696
Kind Code:
A1
Abstract:
The present invention relates to the technical fields of mode recognition, machine learning, multi-task learning, and adversarial attack, and specifically relates to a feature manipulation-based attack and defense method for a continuous learning ability system, aimed at solving the problem that an existing continuous learning-based intelligent system is poor in security and robustness. The method of the present invention comprises: obtaining an image clean sample; extracting a feature of the clean sample; obtaining a target sample, and extracting a feature as a target anchor feature; on the basis of the clean sample feature in combination with the target anchor feature, generating an adversarial sample by means of an attack sample generation algorithm; training an image classification model by means of a continuous learning algorithm, and counting a classification accuracy rate corresponding to the clean sample during C-category task classification and learning; adding, according to a ratio of 1:n, a first matrix as a training sample, and performing retraining; and classifying an image on the basis of the trained image classification model. The present invention improves the security and robustness of the existing continuous learning-based intelligent system.
Inventors:
GUO LIANGXUAN (CN)
CHEN YANG (CN)
YU SHAN (CN)
QU HUI (CN)
HUANG XUHUI (CN)
ZHANG JINPENG (CN)
CHEN YANG (CN)
YU SHAN (CN)
QU HUI (CN)
HUANG XUHUI (CN)
ZHANG JINPENG (CN)
Application Number:
PCT/CN2021/128193
Publication Date:
May 04, 2023
Filing Date:
November 02, 2021
Export Citation:
Assignee:
INST AUTOMATION CAS (CN)
THE SECOND ACAD OF CASIC (CN)
THE SECOND ACAD OF CASIC (CN)
International Classes:
G06N3/04; G06N3/06; G06V10/764
Foreign References:
| CN110334808A | 2019-10-15 | |||
| CN111753881A | 2020-10-09 | |||
| US20210012188A1 | 2021-01-14 |
Other References:
LI XIAOBIN; SHAN LIANLEI; LI MINGLONG; WANG WEIQIANG: "Energy Minimum Regularization in Continual Learning", 2020 25TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR), IEEE, 10 January 2021 (2021-01-10), pages 6404 - 6409, XP033909282, DOI: 10.1109/ICPR48806.2021.9412744
Attorney, Agent or Firm:
HENYOL INTELLECTUAL PROPERTY LAW CORPORATION (CN)
Download PDF: