Highlights
The Binance P2P (peer-to-peer) platform uses advanced large language models (LLM) to monitor transactions for signs of fraudulent behavior and help resolve user appeals.
Some of the most common fraudulent tactics are tricking sellers into releasing their cryptocurrency before receiving actual payment or asking buyers to cancel their order after payment.
By combining AI tools and a diligent Customer Support team, we aim to ensure a safe user experience on Binance P2P.
Check out the work of the unsung heroes working behind the scenes to ensure a secure user experience on Binance P2P.
Binance's P2P platform was launched in late 2018 to facilitate currency exchange transactions between bitcoin and local currencies. While convenient, peer-to-peer (P2P) commerce has its specific risks. Instead of going through a centralized exchange, you rely on another user to fulfill your request to buy or sell cryptocurrency.
What if you are making a transaction with a scammer? Reputable P2P marketplaces, such as Binance P2P, use an escrow service and a strict identity verification process to combat fraudulent activities. But even with all the right protection measures, scammers can find a way to act, and they usually do.
Leveraging artificial intelligence (AI) models, we built a security infrastructure designed to mitigate the specific risks associated with P2P commerce. But before we delve into details, let's take a look at some common scams that traders face when using the “Chat” feature on Binance P2P.
Four common Binance P2P scams
1. Fake Customer Service Representatives
Scammers often impersonate the Binance Customer Support team to trick victims into providing their account or credit card details. They may claim that Binance has already “received payment” and ask the seller to release their escrow cryptocurrencies.
If there is one thing you should remember, it is that our Customer Support team will never, under any circumstances, contact you via Binance P2P chat.
2. Escrow Scam
In this scam, the scammer pretends to be a buyer. During the transaction, the scammer will lie that the Binance P2P escrow service has the fiat payment on hold. It will then claim that Binance will “send” the money once you release your cryptocurrencies.
This is not how the Binance P2P escrow system works. We only temporarily secure sellers' cryptocurrencies, and buyers' fiat payments never go through our escrow service.
3. Threats to call the police
Scammers may claim to have paid after generating an order. If you hesitate, they will pressure you to release your payment by threatening to call the police.
Don't give in to threats on Binance P2P. If you have any dispute or legitimate problem with the counterparty, file an appeal by following the steps in this guide: How to appeal P2P orders in the Binance app.
4. Tricking the buyer into canceling the order after payment
Not all are scams initiated by buyers; Sellers can also implement malicious schemes. After receiving payment, the seller may say that there is a problem with their one-time password (OTP) or payment release and suggest that the buyer cancel the order. The seller will “promise” to refund you the full amount once the order is cancelled.
The seller is just a scammer who never planned to refund the money. Anyone who asks you to cancel an order after you've paid is probably trying to scam you.
Meet the invisible guardians
To protect our users and prevent them from falling victim to the scams we mentioned above, we have our own team of AI heroes working behind the scenes 24/7.
These heroes are specialized AI models trained to detect users acting with bad intentions. The models basically act as gatekeepers, monitoring various phases of the transaction channel with the sole purpose of intercepting fraudulent activities. Below, we'll take a closer look at the models we use and how they work to provide millions of users with a reliable P2P trading experience.
An all-in-one tool: Large Language Models (LLM)
The term large language model (LLM) refers to a general-purpose AI system that is adept at “understanding” and generating human language. LLMs are trained using text data from across the Internet.
Over time, these models can be trained, or fine-tuned, to excel at specific tasks, such as generating original snippets of text or recognizing messages that may indicate the malicious intentions of senders.
How do we use LLMs to train our P2P models?
To tune our models, we expose them to communication data associated with P2P transactions, that is, interactions between merchants. Initially, our models found more examples of general transaction activity than scam-related behavior during the learning process. This posed a major obstacle: how could our models learn how scammers communicate with so few instances to extract data?
We tried several approaches:
Increase the training set of the minority group (scammers samples) by repeating their instances (oversampling) more frequently in the model.
Reduce the number of normal user instances (subsampling).
Modify the importance of each group (altering the relevance of the classes).
None of the three methods were still satisfactory due to the diversity of data from the limited sample size. The most effective approach was to create additional training instances through other LLMs, such as LLaMa 2, OpenAssistant, and Falcon.
We use these LLMs to rephrase existing examples of scammers' communicative behavior, or even invent new examples with similar messages. In this way, we obtained a more balanced set of training material with a satisfactory sample size of scammers for our classification models.
How to understand user intentions
Most of the user interaction on Binance P2P takes place in our built-in chat feature. The content of these conversations can reveal key information about users' intentions. For example, if someone is pretending to be a customer service agent, breaking rules about paying, or needing help finishing an order, they say certain things in chat.
We continually modify our LLMs to identify user intentions in various situations on the P2P platform, as shown in the diagram above. Our models are designed to understand unique situations in our market, as well as to differentiate between suspicious and normal interactions.
Our goal is to prevent scams before they can harm our users. LLMs help us detect suspicious messages before the conversation leads to a transaction. In addition to strengthening security, they routinely help us identify and assist users who need assistance completing a transaction. So far, our AI models have helped us detect and prevent more than 2,000 potential scams and have automatically facilitated the completion of 212,000 appeal chat orders, involving funds totaling more than $28 million.
To better illustrate how our models work, here are two examples of them in action.
Conclusions
At Binance, we invest significant resources in ensuring the security of our users and use the widest range of approaches to achieve that goal, including innovative solutions such as AI-powered tools. In our P2P marketplace, we use large language models to identify users who might be engaging in suspicious behavior. In order to combat the ever-evolving scam industry, our language models are continually retrained to detect the latest tactics and trends.
Our passionate team of Customer Service agents work alongside our AI tools; After all, in some situations nothing can replace a human touch. Together, they ensure that Binance is not only secure but also offers an exceptional user experience, ensuring that all users can rely on the products and features available in the ecosystem.
If you were a victim of a P2P scammer, please file a report with Binance Customer Service by following the steps in this guide: How to report scams to Binance support.
