Xiao Xiao From the Aofei temple
qubits reports | official account QbitAI
Still worried about online shopping may be fake ？
Now? , Ali made his own AI Anti counterfeiter , Let algorithms help crack down on counterfeiting ！
Just show it a picture of the product you want to identify , It can be in Tens of milliseconds In the time of , Judge the authenticity of the goods .
It's like a blink of an eye （ about 0.5 second ）, It's already recognized 10 individual Counterfeit goods .
More Than This , Ali's security team has also disclosed a project containing 100 Data set of ten thousand genuine trademark photos Open Brand, Let more people participate in “AI Counterfeiting ” In action .
that ,AI How to fight against counterfeiting ？
Attack Gao Fang and Shanzhai at the same time
First , We need to know the selling routine of fake goods .
Because there is no qualification to sell authentic products , Fake shops need to use all kinds of means , Escaped the trail ：
firstly ,“ High imitation ” Commodity . This fake will do harm to logo Fuzzy processing 、 Partial occlusion , Even against sample attacks （ you 're right , That's the high technology ）.
Online shopping platform , The fake pictures we see are often like this ：
The appearance of these goods , Often very close to brand goods , However, there is no product guarantee .
second ,“ shanzhai ” Commodity . This fake will do harm to logo Conduct PS、 And use the fake name when registering , To avoid the recognition of the algorithm .
Don't say that , These fake goods , Play “ Fake makeup ” Come on, there's a way ：
Users don't pay attention , It's easy to fall into a trap , However, the speed of platform manual review , And it's really limited .
So , Ali's security team decided , Create a AI“ Anti counterfeiter ”, Fight against counterfeiting faster .
AI Cracking down on fake models , What needs to be solved most , This is it. 3 A question ：
- How to solve the problem of small target recognition ？（ The logo in the picture logo Often very small ）
- How to solve logo Fuzzy deformation coding 、 Even introduce the problem of attack samples ？
- How to accurately identify “ shanzhai ” edition logo？
To solve these problems , Ali's security team designed a project called Brand Net Model of .
According to senior algorithm expert of Ali security Turing Laboratory Chinese flowering crabapple Introduce , The model is divided into three parts ：
Cascade-RPN, This part focuses on solving the problem of detection in commodity pictures logo“ Sometimes big, sometimes small ” The problem of , The accuracy of target detection is improved 、 Small target recall rate .
Soft Mask Attention, This part uses transfer learning , Construct a detection frame （bounding box） Feature to segmentation （segmentation） The mapping space of features , Effectively improve logo smaller 、 Or the detection accuracy in case of deformation .
Feature Representation and Instance Retrieval, This part enables the model to be used in the new logo Elastic expansion when adding , There's no need to retrain the model .
Of course ,Brand Net As AI“ Anti counterfeiter ” One of my tricks , Mainly used for more famous （ There are so many pictures of products ） The brand of , Because they have more pictures of their products , The trained model will work better .
△ The long tail problem of data imbalance
For some products with few photos New brand ,AI“ Anti counterfeiter ” In another way , To improve your identification ability ：
In a semi supervised way , Include 3D Data to enhance , Domain migration （HDA）, Self distillation, etc , To train .
Tests show that , This AI“ Anti counterfeiter ”, stay NVIDIA A variety of graphics cards （T4、P100、V100） It has achieved a very fast detection effect , On average 30~50ms You can identify a fake in a few minutes .
and , The effect of recognition is very good , Recognition accuracy of online pictures , Reached 95%.
Of course , Online shopping platform will not be completely based on AI“ Anti counterfeiter ” The result of our judgment , Punish businesses directly .
stay AI“ Anti counterfeiter ” After finding out the suspected sellers of fake goods , The platform will also start from Product content 、 Sales information 、 Qualifications 、 Consumer feedback And so on , In order to determine whether the business is selling fake goods .
at present , Ali's security team has used technology to crack down on counterfeiting , Assisted the police in several cases , There is a big case of luxury goods and fake goods alone 6 rise , And destroyed the production dens 10 More than a 、 Arrest of criminal suspect 150 More than .
among , Include 2020 year 8 month , Shanghai police made it public “ Shanghai Qingpu produces and sells counterfeit products LV Registered trademark goods case ”, The amount of money involved in this case Over 100 million yuan .
It does sound very coooool.
The coolest thing is , Ali has used it for training logo The data set was made public .
Trained logo The data set is public
According to the Alibaba security team , This public dataset is called Open Brand.
Open Brand It's the biggest luxury in the world logo Data sets , contain 500 Multiple luxury categories 、1000 Multiple subcategories of trademarks , The cumulative amount is more than 100 Thousands of copies trademark logo picture .
Each image , Press COCO structure Marked in detail ：
logo The number of images is also hanging from other data sets ……
According to the Alibaba security team , The photo data in the dataset , All from Taobao 、 Tmall 、1688、Aliexpress, And Google and Baidu search engines .
According to the open data protocol , The photos in these datasets , Only used for Academic research , Not for commercial use .
AI The crackdown on counterfeits , You can also participate in
In order to find a better AI Cracking down on fake models , Ali's security team made a 「 Robust identity detection challenge 」, Bonus 30 ten thousand .
Simply speaking , Ali's experienced security team , Will incarnate “ Fake merchants ”, For the contestants to submit AI Crack down on fake models to make all kinds of tests .
for example , On purpose logo Images are processed in various ways （ Fuzzy 、 Add watermark 、 Generate countermeasure samples ）, have a look AI Can the anti-counterfeiting model be accurately identified .
And the data sets that competitors use to train , It's this big public logo Data sets Open Brand, You can get .
Just take part in the competition , You can get access to College team 、 Safety Daniel A chance to compete .
at present , The rankings are still being updated .
Do you have a heart attack ？
Ali AI Anti counterfeiting competition （ Access to public data sets ）：
Brand Net Address of thesis ：
— End —
qubits QbitAI · Headline signing
Pay attention to our , The first time to learn about cutting-edge technology