Unveiling Alibaba's anti-counterfeiting AI brain: Total data volume: 186 national libraries, 1 AI = 50,000 humans, won the Innovation Award

Li Gen from Xiamen
Quantum Bit Report | Public Account QbitAI

"It is shocking that the United States is so far behind in the fight against counterfeiting!" A month ago, Doug Collins, Republican Senator and vice chairman of the U.S. Judiciary Committee, lamented the problem of counterfeit goods in the United States. His words are still fresh in our minds. Recently, Alibaba's anti-counterfeiting technology has won another high-level domestic technology award.

On August 9, the "China Artificial Intelligence Summit Forum" guided and hosted by the Ministry of Industry and Information Technology, the Ministry of Public Security and the Cyberspace Administration of China was held in Xiamen. After layers of selection, Alibaba's Intellectual Property Protection Technology Brain was named "Artificial Intelligence Innovation Star" by the three ministries.

Intellectual property protection is more commonly known as anti-counterfeiting . Alibaba's intellectual property protection technology brain is an algorithm technology system that aggregates Alibaba's massive online and offline counterfeit feature database and anti-counterfeiting experience accumulated over the past 20 years, with the original "safe AI" brain as its core.

This system operates automatically 24 hours a day, and 96% of suspected infringing links are killed instantly the moment they are posted. In the past three years, Alibaba has used this technology to assist police in 31 provinces and 227 counties across the country in arresting 4,439 counterfeit manufacturing and selling suspects and destroying 4,289 counterfeit manufacturing and selling dens.

From praise from US congressmen to domestic technology awards, it can be said that the time has come and Alibaba's anti-counterfeiting efforts are entering a period of explosive gains. Quantum位 also understands the power and innovation of the security AI behind this, and will reveal them to you one by one.

1 AI brain = 50,000 humans

The technological brain of intellectual property protection is the culmination of sincerity and technological innovation.

This "brain" was developed entirely independently by Alibaba and has been widely used inside and outside the Alibaba ecosystem to detect and combat counterfeit goods, copycat products, infringement and piracy.

Moreover, as an AI on the job, it has strengths in all aspects. Not only is it better than humans, but it also has many abilities that humans cannot do.

Alibaba announced a series of amazing data performance:

If it takes 10 seconds for a human to check the information of a picture, then 50,000 people working at the same time can barely catch up with the speed of the "Intellectual Property Protection Technology Brain" scanning and identifying pictures. The number of new products released on the Taobao Tmall platform every day is in the tens of millions. If it is manually inspected, it will take 138,889 people working for one day to complete this task.

Xue Hui, head of Alibaba Security Turing Lab, revealed that the total amount of sample data of this system is equivalent to the collections of 186 National Libraries of China. The cumulative sample volume of anti-counterfeiting pictures alone exceeds 13.7 billion. When printed on 0.3 mm standard photo paper and stacked, it is as high as 4,110 kilometers, which is 4,964 times the height of the world's tallest building, the Burj Khalifa, and approximately equal to 464 Mount Everests.

Behind the amazing effect is the "Security AI" risk control system forged by Alibaba's continuously evolving technological strength in security scenarios.

How does secure AI serve intellectual property protection?

The so-called security AI refers to AI evolved in security scenarios, which has stronger and more diversified capabilities. Different from the current concept of general AI, Alibaba Security focuses more on the vertical and in-depth accumulation of technology in the field of business security, and develops a new generation of AI that adapts to more security scenarios, so that traditional security problems can find new AI solutions. Alibaba proposed this new concept at the beginning of this year, saying that it will become the core solution to future network security problems.

It is reported that the secret of Alibaba's intellectual property technology brain lies in four core intelligent engines, covering hundreds of Alibaba's independently developed security AI technologies. Specifically, they involve the perception engine, cognitive engine, decision engine and computing engine.

In fact, they are also the four indispensable engines of general artificial intelligence (AGI), but Alibaba’s vertical service is to combat counterfeiting.

The macro-architecture context can also be found in the project application instructions.

Perception Engine: The core technologies are computer vision and speech recognition technologies. It is the sensory system of the entire system, the basis for object recognition, and the first step in forming a series of subsequent processing actions.

Including biometric technologies such as face recognition, voiceprint recognition, and liveness detection for identity authentication in the store opening process; document recognition and tampering detection technologies for store qualification verification; trademark detection, object detection, and optical character recognition technologies for online product identification; advertising image analysis and live video monitoring technologies for marketing detection; as well as multimedia retrieval and image watermarking technologies for the protection of original works.

It is the "sensory system"'s precise voice, image, audio, and video perception capabilities that enable the entire system to accurately identify objects that need protection and defense among hundreds of millions of complex data and proceed to the next step of processing.

Cognitive engine: The core technology is natural language processing technology.

The cognitive engine is the system's "translator", allowing the machine to understand the meaning of text, the meaning of human voices, and the nature of people's actions, so as to judge good intentions, bad intentions, danger and safety.

Specifically, it includes knowledge graph technology for building massive structured and unstructured commodity data; multilingual analysis and machine translation technology for overseas e-commerce, semantic analysis technology for identifying external feedback evaluations and public opinions; and technology for integrating multimodal information for comprehensive recognition and understanding of commodities.

Third, decision-making engine: core technologies include deep learning and reinforcement learning, which are used to solve complex decision-making problems in reality.

It is the "strategist" of the system, determining how the system should act and how to make the best response more reasonably and effectively. For example, when faced with real-time gaming among merchants, changes in information content, and malicious attacks from black industries, it can make more favorable decisions from a global perspective.

Finally, there is the computing engine , which includes a real-time indicator computing system, a distributed heterogeneous computing system, and a large-scale graph neural network system.

This is the engine of the anti-counterfeiting AI brain. Its powerful performance ensures that the system can remain calm, accurate and efficient in the face of thousands of concurrent and billions of data.

Behind the computing engine is Alibaba Cloud's machine learning platform - PAI 3.0.

It can support concurrent training of thousands of workers for a single task and 5k+ ultra-large-scale heterogeneous computing clusters, ensuring round-the-clock monitoring of billions of commodity abnormalities and all-round supervision of operators' behavior.

One minute on stage, 20 years of technical skills

It is worth mentioning that although the four major engine architectures are macro, every technological accumulation behind them is the result of engineers’ hard work day and night.

It is worth mentioning that the anti-counterfeiting AI brain is one of the application scenarios of Alibaba Security's "Security AI". Alibaba continues to evolve the power of AI in high-risk and strong confrontation scenarios, and continuously applies new technologies such as small sample learning, multimodal, and self-supervised learning in more than 100 scenarios such as content security, new retail security, and transaction security. It embodies Alibaba Security Turing Lab's more than 10 years of technical accumulation in the field of AI, and achieves the goal of dripping water wears away a stone.

It is not difficult to discover the strength of Alibaba Security Turing Lab from its recent paper presented at the top artificial intelligence conference.

This is a paper published by Alibaba at ECCV Workshop 2018, which mainly shares its technical progress in the field of video analysis .

This is also Alibaba’s secret to identifying and combating counterfeits in videos and protecting originality.

The current video analysis method commonly used in the industry often pre-trains the CNN network to extract feature classification, and then uses recurrent neural networks (RNN, LSTM) for sequence modeling.

However, the feature sequence of a video is generally long and contains a multi-level structure (hierarchical data structure), that is, a video contains frames, shots, scenes, events, etc.

Moreover, the relationship between frames and shots is very complex, not just the sequential relationship between previous and next frames. Through general sequence modeling methods, RNN cannot express such complex relationships, and the modeling effect is poor.

Therefore, the Alibaba research team used the deep convolutional graph neural network (DCGN) to perform multi-level modeling of video frames, shots, and events, gradually abstracting from the frame level, shot level, to the video level, thereby obtaining a global expression of the video and then performing classification:

After the final method was verified on the youtube8m dataset, the effect was improved compared with other classic sequence modeling methods.

Paper portal: https://arxiv.org/abs/1906.00377

There are also advances in the field of natural language processing , which is also a key technology for the scientific and technological brain of intellectual property protection.

The article by Alibaba Security Turing Lab was selected for IJCAI 2019. The theme is object-oriented sentiment analysis. The main goal is to discover the objects of comments and determine the polarity of emotional expression.

Generally speaking, many clues about product quality descriptions are hidden in users' comments on the products, but it is difficult to find problems in the products themselves. This is also a scenario where NLP technology can be used.

In layman's terms, the method proposed in the paper uses global information and context to identify emotional objects, rather than predicting the sequence label corresponding to each word.

This method is based on the word chunk method and proposes a simpler and more efficient joint model to simultaneously extract the emotion expression object and determine its emotion polarity.

In terms of specific steps, we first represent all candidate word blocks in the comment sentence with vectors, and then propose a word block-based attention mechanism to predict the labels and polarities corresponding to the word blocks.

Finally, after evaluation and comparison with public datasets, it is proved that the performance is better than the existing methods.

Another aspect that reflects the challenges of the scenario and the level of technology is the adversarial problem that is unique to security scenarios .

Counterfeit sellers often change the "title" and "description" to avoid identification by traditional rules and models, but this is not completely untraceable.

Obfuscated language is a technique used to evade detection in adversarial communication scenarios.

Adversarial communication scenarios include the dissemination of sensitive information, expression of negative emotions, planning of covert actions, and illegal transactions. Obfuscated language is usually achieved by replacing variant words in the original text.

When regulators identify such texts, they need to scan and filter them based on a set of keywords. Although some semantic expansion techniques have been introduced, the accuracy and recall of identifying such texts are very limited due to the ambiguity and boundless variation in the text.

The paper published by Alibaba at WWW 2019 focused on disclosing the core progress in this direction.

The main idea of ​​this paper is to transform obfuscated language recognition into a text matching task, that is, whether each piece of information to be detected matches a scanned keyword, and at the same time integrates the textual representation and visual representation of the text information.

The visual representation here refers to the visual effect of the text itself, rather than the pictures in the information. This is mainly because when performing text mutation obfuscation, some characters that look similar are often used for replacement. This mutation causes the obfuscated content to have no semantic connection with the original content, but it can be connected in terms of visual effects.

Alibaba's model uses BiLSTM to represent text features and represents visual features through template matching. Through multimodal integration, it can show higher accuracy and recall rate than traditional methods.

Another is the multimodal task solution that combines image, video and natural language processing .

The paper was published at ICASSP 2019, and the tasks listed in it are also quite interesting: through a text description, the source image is automatically edited to conform to the given text description, thereby simplifying the image editing process, which is a text-based image editing method. The products on e-commerce websites are a mixture of text and pictures, so it can be expected that this technology can enhance the understanding of the connotation of the product and help to detect fake and inferior products.

Interested friends can move to the portal: http://arxiv.org/abs/1903.07499

Finally, let me introduce a sexy technological advancement in the industry that has made a significant contribution to anti-counterfeiting AI: small sample learning .

The most troublesome point in the counterfeit problem (security issue) is the lack of sufficient training samples for emerging risks, which makes many excellent machine learning algorithms hesitate.

At CVPR 2018, the Alibaba Security Turing Lab team presented a solution to the industry-leading challenge of "zero-sample video retrieval".

Video retrieval,usually needs to extract cross-modal correlations between text and,video, based on content matching.

But what makes Alibaba's approach different is that it proposes a content-independent method that encodes video and text into dense representations of their respective modalities through a dual deep encoding network.

Furthermore, the dual encoding concept is simple, effective, and can be learned end-to-end.

Experiments on three benchmark datasets, MSR-VTT, TRECVID2016, and 2017, prove that the zero-sample video retrieval method proposed by Alibaba has reached the best level at present.

Paper portal: http://arxiv.org/abs/1809.06181

At the recent top machine learning conference IJCAI-2019, Alibaba Security successfully held the first AAAC competition (Alibaba Adversarial AI Challenge) and AIBS seminar (Artificial Intelligence for Business Security), aiming to explore how to solve the security issues of AI models when facing adversarial attacks. The competition and conference attracted more than 2,000 teams from 24 countries and regions to participate, and many new ideas and methods emerged in the process, greatly promoting the development of this field.

"The development trend of AI is certain, but AI cannot be applied mechanically to solve security problems. AI technology needs to be upgraded according to actual scenarios," Xue Hui pointed out in a public interview at the beginning of this year. Security will become the biggest challenge for future AI development, and "Secure AI" will become a new solution to future network security issues and will also usher in an explosive period in 2019.

Today, the successful practice of Alibaba's "Intellectual Property Protection Technology Brain" has been continuously recognized by the industry and even the world, which confirms this assertion.

Alibaba's anti-counterfeiting AI uses technology to solve social problems

The reasons behind the success of the anti-counterfeiting AI brain are not difficult to analyze. This will also be a guarantee for Alibaba's various businesses to prosper in the AI ​​era.

Fighting counterfeiting is a social problem caused by a combination of factors, and it is not easy to achieve it using technologies such as AI. Without comparison, it may be difficult to see sincerity.

For example, Amazon, the American e-commerce giant, actually launched a new anti-counterfeiting project called "Project Zero" this year, which aims to combat and eliminate counterfeit goods by cooperating with brands.

However, in terms of specific technical mechanisms, based on the current disclosures, it is slightly inferior to Alibaba's anti-counterfeiting AI brain.

Because of Amazon's anti-counterfeiting AI, it also requires partner brands to provide logos, trademarks and other information, and even allow brands to use tools to mark and ban counterfeits.

It still relies more on supervised learning, and the degree of automation and general AI capabilities are relatively insufficient. Compared with Alibaba's unsupervised learning, small data learning and systematic brain, the technical capabilities and challenges are clearly different.

It is no wonder that after investigating e-commerce platforms including Amazon, eBay, and Alibaba, Doug Collins, vice chairman of the U.S. House of Representatives Judiciary Committee, commented: "Alibaba's anti-counterfeiting policies and programs are much more effective than any of its American counterparts."

Ali's consistent way

Finally, Alibaba’s way of doing things is still worth paying attention to.

From a macro perspective, intellectual property protection of the technological brain is still another successful Alibaba-style innovation.

The source of the problem of fighting counterfeiting lies in the stage of social development and the bad roots of human nature. The challenges we face are essentially the same as those faced by e-commerce, payment, logistics, computing and independent chips, and the difficulties are no different.

However, Alibaba's approach remains consistent. Its AI brain system, which was created to combat counterfeiting, is still a continuation of the Taobao Tmall, Alipay, Alibaba Cloud, Cainiao, and Pingtou Ge models:

Technology-driven, building a system platform to fundamentally solve the problem. Moreover, the great heroes will also benefit the world in the future, empowering all walks of life and even all countries, and everyone will benefit from it.

Will there be a day when there are no more fakes in the world? Maybe we can still have dreams.

After all, AI can do it, Ali is doing it…

You are also welcome to tell us what other AI technologies can be used to fight counterfeiting^_^