Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
ATTACK METHOD AND DEVICE FOR EVALUATING ROBUSTNESS OF OBJECT DETECTION MODEL
Document Type and Number:
WIPO Patent Application WO/2024/027068
Kind Code:
A1
Abstract:
The present application provides an attack method and device for evaluating robustness of an object detection model. The method comprises: acquiring at least one object picture inputted into an object detection model; using an initial importance matrix to calculate key pixel points in the at least one object picture; selecting the key pixel points as disturbance objects according to importance; performing iterative attack on the key pixel points until the attack is successful and the number of iterations is reached; when the attack is successful and the number of iterations is not reached, on the basis of disturbance amplitudes, recovering a first preset number of key pixel points having disturbance amplitudes smaller than a preset value; when the attack fails and the number of disturbance iterations is reached, adding a second preset number of new key pixel points; and storing an object picture having the minimum number of disturbed key pixel points and the minimum number of detected object instances in an attack process. According to the present application, disturbed key pixel points are dynamically added and recovered in an attack process, thereby facilitating finding a better disturbed key pixel point combination.

Inventors:
LV JIEYIN (CN)
DAI TAO (CN)
LIU HAO (CN)
ZHOU SHOUQIN (CN)
Application Number:
PCT/CN2022/137578
Publication Date:
February 08, 2024
Filing Date:
December 08, 2022
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
SHENZHEN CIMC INTELLIGENT TECH CO LTD (CN)
SHENZHEN CIMC HIGH TECH CO LTD (CN)
International Classes:
G06V10/82
Domestic Patent References:
WO2021257817A12021-12-23
Foreign References:
CN114998707A2022-09-02
CN113780123A2021-12-10
CN114220097A2022-03-22
CN114332569A2022-04-12
US20200265271A12020-08-20
Attorney, Agent or Firm:
P. C. & ASSOCIATES (CN)
Download PDF: