AI Visual Anti‑Fraud Model Battles QR Code Abuse in the Beverage Industry

Unexpectedly, the “Open the cap for a prize” activity was targeted by black‑gray industry and coupon‑hunting groups, leading to multiple consumer complaints online.

“The QR code looks clear, but no matter how I scan it, it never works. I tried several phones and still can’t read it; I suspect the QR code is invalid.”

This is not an isolated incident. Consumers who encounter invalid QR codes often complain about false advertising or suspect that the beverage has been opened, raising food‑safety concerns.

Behind this lies a massive black‑gray industry chain.

QR codes can leak at production, transmission, printing, or activation stages; before a consumer scans the bottle cap, coupon‑hunting groups have already purchased batches of QR codes through illegal channels. In one Guangdong beverage factory, the fraudsters once siphoned tens of millions of yuan.

The beverage company tried to solve the leakage by installing cameras in production and printing, tightening employee management, limiting scan counts, blacklisting abnormal users, and prohibiting rescans…

The company spent 200 million yuan and deployed hundreds of staff to build a risk‑control system to stop collective theft by the black‑gray industry, but the results were unsatisfactory.

1. Fighting a black‑gray industry worth hundreds of billions is hidden in a tiny bottle cap

In summer 2021, the beverage maker approached Ant Group’s “Xiaoling” (pseudonym) and described losses of over a hundred million yuan. “You may be our last lifeline!” they pleaded, hoping to return the red packets to genuine consumers.

To understand the black‑gray operation, Xiaoling went undercover, joining over 500 coupon‑hunting chat groups, paying for their “scanning robot” courses, attending events, and even purchasing the automated QR‑code redemption robots.

Analysis revealed the core problem: the location of the code. Fraudsters scan codes displayed on screens or printed on paper, while genuine consumers scan codes printed on bottle caps or scratch cards.

Xiaoling said she needed a pair of eyes—once you can “see” where the code is, distinguishing real from fake consumers becomes trivial.

Those eyes are an “AI visual anti‑fraud model” developed by Ant’s AI vision engineers Guo Hai and his team.

The team, which publishes five or six papers annually in top journals and conferences, had built an image perception platform three years earlier, aggregating advanced image‑processing algorithms.

“Now it’s optimized so that a single button triggers the function,” Guo Hai demonstrated, assembling the first AI visual anti‑fraud model in half a day.

Ant has long accumulated expertise in visual algorithms. The platform has undergone multiple upgrades, handling tasks from document authentication to pet‑nose identification and remote sensing of tree growth. The team won first place in the ICCV 2021 OVIS competition.

AI visual algorithm detecting dog nose feature points

Remote sensing and AI data analysis reveal distant tree growth

2. In less than 0.1 second, AI’s “eyes” can identify genuine consumers

Guo Hai opened the folder Xiaoling sent, filled with bottle caps bearing QR codes collected from restaurants, street stalls, and even from family members during rainstorms.

After feeding 3,000 cap samples into the model, in under 0.1 second the AI distinguished redeemable genuine scans from suspicious ones, prompting re‑scans for the latter.

“That’s not enough,” Xiaoling said. Changing the shooting angle or hiding the code behind a phone or computer frame could fool the model.

Therefore they introduced unsupervised anomaly detection, leveraging environmental cues for decision‑making, and fed 12,000 black‑gray QR codes as negative examples for iterative upgrades.

Subsequently, Xiaoling and the red‑blue confrontation team simulated black‑gray attacks while Guo Hai’s team continuously defended and upgraded the model.

The system went live in October 2021, and the flood of consumer complaints vanished quickly.

But this is only the beginning.

Guo Hai iterating the AI visual anti‑fraud model

3. In August, the model iterated over 100 versions, driven by countless late‑night battles against black‑gray attacks

Facing billions of profit, the fraudsters also evolve.

Xiaoling observed that fraudsters progressed from scanning screens to stitching QR code edges with socks, then PS‑ing QR codes onto caps.

Initially Guo Hai could spot issues with the naked eye; later he had to magnify doubtful areas to find subtle differences before feeding them to AI.

Black‑gray tricks deceive algorithms but are intercepted by AI

Attacks usually occur at night; a few hours of fraud can cause tens of thousands of yuan loss, while Ant’s team stays online to respond to every late‑night query.

During labeling, one person annotates seven or eight thousand images per day.

Guo Hai asked, “How many more?”

“Four thousand,” was the answer.

“Give me two thousand then.”

In the labeling process, they discuss how to detect Moiré patterns when high‑end phones hide them, and how to analyze QR code surroundings for tampering.

Xiaoling said Guo Hai often brings a glimmer of hope; whenever a new problem appears, the team quickly proposes new ideas.

Through prolonged offense‑defense cycles, the AI visual algorithm grew stronger; new fraud techniques are promptly blocked, and image‑forgery cues such as Moiré patterns and PS artifacts are incorporated into detection.

From last year to now, the AI visual anti‑fraud model has iterated over 100 versions in eight months, achieving 99.3 % accuracy and becoming a comprehensive risk‑control system that combines multimodal, multidimensional analysis; Xiaoling and Guo Hai filed two patents.

“Black‑listing fraudsters was era 1.0, rule‑based judgment was 2.0, and image‑forgery‑based risk control is 3.0,” Xiaoling summarized.

Now the platform aims for full automation, allowing AI to select problematic images and learn without constant human training.

With the technology mature, red packets return to consumers, and the system can also handle QR codes linked to pornography, gambling, or scams by extracting surrounding text and images to protect users.

The same capability can be applied to satellite remote‑sensing imagery. When a satellite captures crop photos, the system instantly analyses species, vigor, and yield forecasts.

NetBank “Sparrow” satellite remote‑sensing risk‑control system

Can estimate crop value and help farmers obtain satellite loans

A “lender” that knows terrain, precipitation, accumulated temperature, and historical yields can assess crops, performing tasks traditional loan officers cannot, making financing easier for farmers.

Beyond that, Xiaoling and Guo Hai look forward to applying AI visual algorithms to more domains.