Training data coverage includes real-world fraud scenarios, with a high protection rate against deepfake threats and fairness across demographics.

Key applications include high-detection-rate account protection against deepfakes, enhanced eKYC compliance across markets, and advancing financial inclusion in emerging markets.

Ant International operates a broad ecosystem serving over 150 million merchants and 1.8 billion user accounts, with AI Security initiatives like AI SHIELD cutting AI service risk by 90% and Alipay+ EasySafePay 360 reducing account takeover incidents by 90%.

AI fairness is framed as strengthening security and reliability in identity verification, with a >99.8% deepfake detection rate aimed at protecting accounts in 200 markets.

Headquartered in Singapore, the company runs across Asia, Europe, the Middle East, and Latin America, offering a unified techfin platform for digital payments and financial services.

Ant International plans to integrate the winning security technology into its payment and financial services to strengthen security, eKYC compliance, and deliver a consistent user experience globally.

The winning approach uses a Mixture of Experts architecture with an internal bias-detection mechanism to train sub-networks for deepfake detection while ignoring sensitive demographic traits, achieving fairness across gender, age, and skin tone.

The MoE-based model combines internal bias detection to train sub-networks tasked with deepfake detection and neglects sensitive demographic traits, improving fairness and accuracy.

The competition tested models on 1.2 million AI-generated human face images across diverse demographics to emphasize fairness across gender, age, and skin tone.

The model is trained on globally representative data, including real-world payment fraud scenarios, and delivers robust, fair performance at scale.

This MoE architecture with bias-detection delivers fair and accurate performance across demographics while enhancing deepfake detection.

Ant International reports a detector performance exceeding 99.8% in identifying deepfakes, providing robust protection against account takeovers across 200 markets.