Binance P2Pâs Invisible Guardians: Using Computer Vision to Detect Fraud
Main Takeaways
Utilizing artificial intelligence (AI)-based computer vision (CV) models for assessing proof of payments, we offer the users of Binanceâs P2P platform an added layer of protection against fraudulent activity.
Our commitment to enhancing user experience manifests in maintaining a simple yet robust security framework within our P2P platform. We believe that even with the use of cutting-edge technologies like AI, security should not come at the cost of complexity.Â
A mindful approach to carrying out peer-to-peer transactions is the first step toward ensuring the safety of your assets. Always exercise vigilance upon receiving any proof of payment. Confirm that the actual amount received matches the declared transaction value.
Deception of Perception: The Threat of POP ManipulationÂ
In the previous blog post of this series, Binance P2Pâs Invisible Guardians: Using AI to Safeguard Crypto Users, we discussed how we use large language models (LLMs) to provide real-time oversight over P2P usersâ communications to help detect potentially dubious transactions. Now, we shift our focus to another aspect of the challenge of leveraging cutting-edge tech to protect P2P users and look into the application of computer vision (CV) models to detect scammers.
In peer-to-peer transactions, the role of valid proof of payment (POP) is fundamental. It's paramount for users to employ due diligence to make sure that they have received the accurate trading sum. Nevertheless, editing tools, even the simplest ones, are readily available for unscrupulous individuals to fabricate seemingly authentic POPs. This poses a real challenge for those on the receiving end to identify counterfeit transactions. Scammers may also make the trading process even more intimidating by using various social engineering tactics such as rushing or confusing the victim.Â
Overall, more than 250,000 images are sent on our platforms daily, not limited to POP images. The variations in these POPs are manifold â the images differ in format, design, and source, which adds another layer of complexity to the task of detecting manipulations.
While the threat of fraudulent proof of payment (POPs) persists, our users are not left to combat these deceptive POPs on their own. Equipped with advanced AI algorithms, our platform scrutinizes all images submitted during transactions, ensuring their authenticity and legitimacy in real time. These robust models are adept at pinpointing and flagging any signs of image tampering or manipulation. Such rigorous safeguards are implemented to secure the integrity of our users' transactions, preemptively warding off fraudulent attempts and thereby mitigating the risk of potential user losses.
Unmasking Deception: Leveraging Computer Vision to Spot Fake Payment Proofs
Computer vision (CV), a field of artificial intelligence where computers are trained to interpret and understand the visual world, offers one of the most robust solutions available today to the issue of forged proof of payment.Â
By using techniques such as optical character recognition (OCR), image classification, object detection, and digital image processing, CV models can be trained to differentiate between genuine and fraudulent proof of payment.Â
These models can analyze a photograph or a screenshot, pixel by pixel, detecting inconsistencies and anomalies that would typically be imperceptible to a human eye. Whether it's a slightly skewed color scheme, distorted logo, or subtle font changes, the model can identify these discrepancies accurately.Â
In conjunction with using the CV models, we also conduct additional cross-verification with the imageâs metadata, as well as other parameters such as trading patterns, to obtain a comprehensive view of the situation.
In the following sections, we will take a closer look at some of the checks we have in place to ensure the legitimacy of POP.
Fig. 1: [Activity] Are you able to identify what information has been tampered with?
Optical Character Recognition (OCR)
A key weapon in our arsenal against scammers is a technique called optical character recognition (OCR). OCR models aid us with establishing two essential pieces of information:Â
The textual content that has been identified and transcribed from the image.
The position of these identified and transcribed pieces of text within the image.
In the world of online transactions, advanced editing tools are readily available. Scammers often exploit these tools to manipulate parts of an authentic POP. Some of the commonly manipulated areas of a legitimate POP include:
Transaction date and time. This field represents the exact date and time when the transaction was processed. Scammers may change this information to reflect a timeline favorable for their deceit. OCR can help verify the accuracy of this data by cross-referencing it with the time stamp of the actual transaction.
Transaction ID. This is a unique identifier for each transaction. Any inconsistency in this field is a telltale sign of tampering. OCR helps in verifying the transaction ID by cross-referencing it with past transactions and scanning for any reused ID.
Transaction amount. Perhaps the most commonly manipulated field. Altering information in this field can lead to discrepancies between the actual and transmitted values. Here, the OCR system's ability to accurately identify numbers plays a critical role in preventing fraud.
Name. The payer's and payee's names are also tampered with quite frequently. OCR models can aid in cross-verifying the extracted information from the name field by checking them against the known credentials of the involved parties.
Fig. 2. Highlights of the tampered information. In reality, the edits would be less noticeable to the untrained eye.
Digital Image Processing
Sometimes, the smallest details tell the largest tales. In the context of image manipulation, any alterations made to an image, no matter how minute, can leave traces or anomalies that are picked up by our models. These faint signals often serve as tangible evidence of tampering. Areas that are most vulnerable to this form of manipulation, as outlined in Figure 1, are more likely to reveal if an image has been altered.Â
To better explain the challenging task of spotting these anomalies, we have identified some common types of manipulation traces:
Unnatural text formatting or alignment. Most noticeable is the text that is tilted, misaligned, or features a starkly different font from the rest of the image. This is usually a fairly apparent sign of tampering.
Subtle background color discrepancies. More subtle are faint differences in the background color that might escape a casual observer but not our comprehensive model.Â
Pixel-level distortions. The least noticeable anomaly is noises or inconsistencies at the pixel level, predominantly around the manipulated fields.Â
An invaluable tool employed to identify such traces of manipulation is Error-Level Analysis (ELA). ELA inspects the compression level across an image. In essence, it identifies areas of an image that show different levels of compression than those in the surrounding area, which can signify that they have undergone recent editing. An original and untouched image will have relatively consistent error levels throughout. In contrast, adjusted parts of the image will display a brighter or more vivid ELA result.
Fig. 3. Example of the resultant image after applying ELA, which provides some clues about the areas that may have been tampered with.
Our models are constantly trained and updated to adapt to the ever-evolving techniques employed by scammers, ensuring the platformâs integrity and preserving the trust bestowed upon us by our users.
Similar Image Search
While we've addressed instances where images are manipulated, there are other equally devious tactics employed by scammers. Another commonly used scamming method is to repeatedly use identical or slightly modified legitimate POP for multiple transactions. If a victim doesn't verify the transaction by checking their bank account, they might unknowingly fall for this scam.
Given the colossal number of transactions and their associated images, the task of scanning and comparing each image is no mean feat. Its real-time implementation is resource-intensive and thereby practically challenging.
To rise to this challenge, we employ an image encoder that condenses images into smaller but vital data abstracts. These snippets are stored in our robust vector database, enabling the algorithm to conduct near real-time scans for similar images. This systematic approach has proven to be extremely effective, allowing us to thwart hundreds of scam attempts daily. The utilization of a similar image search algorithm is yet another testament to our commitment to proactively ensuring the safety of transactions on our P2P platform.
Fig. 4. Near real-time vector search pipeline
Closing ThoughtsÂ
In the world of peer-to-peer crypto transactions, taking on the challenge of fraud is increasingly vital. We stand at the frontlines, constantly refining and leveraging advanced technological solutions to fortify our defenses against fraudsters that target our platform and users.
Our use of AI exemplifies our proactive approach to securing the integrity of every image submitted in transactions. Using real-time surveillance and analysis, these powerful AI models can accurately and quickly identify image manipulation attempts. The effectiveness of these measures is remarkable, particularly given the vast volume and diversity of images exchanged daily across our platform.
However, we also believe that security should not come at the expense of user experience. We remain committed to ensuring that all users enjoy a straightforward, seamless experience on our P2P platform, without worrying about the integrity of their trades. We champion the principle that secure transactions and user-friendly navigation are not mutually exclusive, but complementary scopes in the path to a thriving digital trading environment.
The battle against fraudulent P2P transactions does not stop at the deployment of advanced technology. It also requires the vigilance and engagement of our user community. By combining the formidable technological arsenal at our disposal and the user community's active participation, we can offer a secure and reliable marketplace.
Note
For cases where our models identify highly suspicious POP, you may see the following cautionary message appearing in your chat box:
Log into your payment account and verify the correct payment has been received. Otherwise, DO NOT release before checking.
Be sure to check your account!
If you have fallen victim to a P2P scam, please file a report to Binance Support by following the steps in this guide: How to Report Scams on Binance Support.