How to Fight Algorithmic Bias in Digital Payments (Award-Winning Solution)

Ant International, a leading digital payments and fintech company, has claimed victory at the prestigious NeurIPS Competition on Fairness in AI Face Detection. The win underscores the company's commitment to building secure, inclusive financial services — a mission that is increasingly critical as deepfake technologies become more sophisticated and widespread.

📌 Why it matters: Algorithmic bias in facial recognition is no longer just an ethical concern — it is a direct security vulnerability that puts millions of people at risk of being denied essential financial services.

🔍 The Problem: Bias in AI Facial Recognition

The rapid adoption of facial recognition across industries has exposed a troubling pattern of algorithmic bias. Research by the National Institute of Standards and Technology (NIST) reveals that many widely used facial recognition algorithms show significantly higher error rates when analysing the faces of women and people of colour.

This disparity is rooted in two core issues:

• ⚠️ Lack of diversity in AI training datasets
• ⚠️ Homogeneous demographics among those building and controlling mainstream AI platforms

The consequences are severe: biased algorithms can deny financial services to large segments of the population and introduce critical weaknesses into security protocols.

🏆 The Competition: NeurIPS Fairness Challenge

The NeurIPS Competition was held alongside the Conference on Neural Information Processing Systems — one of the world's most respected AI conferences. Participants were challenged to build AI models capable of achieving both high performance and demographic fairness across three key dimensions:

Ant International's team outperformed over 2,100 submissions from 162 teams worldwide. The task: accurately detect 1.2 million AI-generated face images carefully selected to represent a broad spectrum of demographic groups.

🤖 The Winning Approach: MoE Architecture + Bias Detection

Ant International's award-winning AI model is built on a Mixture of Experts (MoE) architecture combined with a custom bias-detection mechanism. The system works by training two competing neural networks simultaneously:

• 🧠 Network 1: Focused on identifying deepfake content
• 🧠 Network 2: Designed to challenge Network 1 — forcing it to ignore demographic characteristics when making decisions

This adversarial dynamic ensures the model learns to detect genuine signs of manipulation rather than inadvertently relying on demographic patterns. The training incorporated a globally representative dataset and real-world payment fraud scenarios to ensure robust performance at scale.

💬 "A biased AI system is inherently an insecure one. Our model's fairness isn't just a matter of ethics; it's fundamental to preventing exploitation from deepfakes and ensuring reliable identity verification for every user."
— Dr. Tianyi Zhang, General Manager of Risk Management & Cybersecurity, Ant International

💵 Real-World Application: Payments & eKYC

The technology behind the winning model is now being integrated directly into Ant's payment and financial services. Key performance highlights include:

This is particularly significant in emerging markets, where algorithmic bias has historically hampered financial inclusion efforts, preventing underserved populations from accessing basic financial services.

🛡️ AI SHIELD: Ant's Broader Security Framework

Ant International serves over 150 million merchants and 1.8 billion user accounts through services including Alipay+, Antom, Bettr, and WorldFirst. AI security is a declared pillar of its operations.

Central to this is AI SHIELD — a comprehensive risk-management framework built on AI Security Docker, designed to mitigate vulnerabilities in AI services such as unauthorised access and data leakage. AI SHIELD underpins a suite of solutions that provide broad protection for financial transactions, including:

• 🔒 Defences against deepfake attacks and fraud
• 🔒 Alipay+ EasySafePay 360 — which has reduced account takeover incidents in digital wallet payments by 90%

📷 Image source: "abstract art of a beautiful portrait, solid shapes, geometric shapes, neotokyo colors, muted colors, pixar, artstation" — public domain.