INTERLAYER PARSING-BASED INPUT INSTANCE VERFICATION METHOD FOR NEURAL NETWORK MODEL

WIPO Patent Application WO/2021/027052

A1

Disclosed is an interlayer parsing-based input instance verification method for a neural network model, the method comprising: providing a neural network model and a training data set thereof, extracting intermediate information and generating a sub-model corresponding to each corresponding layer; for any input instance to be verified, inputting same into a sub-model to acquire a total behavior profile after interlayer parsing; analyzing the interlayer parsing profile of the input instance; verifying whether the input instance is valid; and providing a confidence score regarding the validity. In the present invention, the validity of a given input instance is analyzed by, on the basis of an interlayer parsing means inside of a training module, using the behavior of the input instance during the parsing of each layer of the model, which may avoid the disadvantage of verification consuming a lot of time which exists for the verification means in the prior art in which many different models must assist on another, and which may core accurately perform input verification, thus helping to distinguish the validity of an input instance of a given neural network, thereby improving the accuracy and security of the neural network during actual use.

XU JINGWEI (CN)
WANG HUIYAN (CN)
XU CHANG (CN)
MA XIAOXING (CN)
LYU JIAN (CN)

PCT/CN2019/111612

February 18, 2021

October 17, 2019

Click for automatic bibliography generation  

UNIV NANJING (CN)

G06N3/04

WO2019098418A1

2019-05-23

CN109344806A	2019-02-15
CN105184678A	2015-12-23
US20190156933A1	2019-05-23

NANJING LE YU ZHI XING PATENT AGENCY FIRM (GERERAL PARTHNERSHIP) (CN)

View/Download PDF      PDF Help

Previous Patent: ILLUMINATION DEVICE HAVING SILENT HIGH-VOLUME AIR DUCT STRUCTURE

Next Patent: TFT ARRAY SUBSTRATE AND DISPLAY PANEL COMPRISING SAME