AutoPilot: AI‑Driven Unmanned Risk Control in Alipay’s AlphaRisk System

Background

With the rise of artificial intelligence, Alipay’s risk engine has transitioned from the CTU era to the AlphaRisk era, introducing AI‑driven, autonomous risk‑control capabilities. AutoPilot, one of AlphaRisk’s four core modules, aims to deliver precise, personalized user authentication recommendations.

Core Concepts

1) User Segmentation – Combines decision‑tree algorithms with the Delphi method to create stable, business‑meaningful clusters that consider both data‑driven stability and specific risk probabilities.

2) Multi‑Objective Optimization – Seeks the optimal balance among disturbance rate, coverage rate, failure rate, and limit‑rate, a classic multi‑objective problem solved by approximating the Pareto‑optimal front.

Because real‑world multi‑objective problems involve conflicting goals and a vast solution space, a set of Pareto‑optimal solutions is generated, after which a single solution is selected using one of three strategies: decision after search, search after decision, or simultaneous search and decision. Evolutionary algorithms are well‑suited for this task.

Evolutionary Algorithm Overview

Maintain a population of candidate solutions.

Evaluate each individual’s fitness.

Perform selection, keeping high‑quality individuals.

Apply crossover and mutation to produce the next generation.

The goal is to approach the Pareto front while preserving diversity across the population.

Specific Algorithm: Random Weight‑Based Genetic Algorithm (RWGA) with Niche Method

step1: Generate an initial random population E; step2: Assign fitness to each individual by aggregating multi‑objective functions with random weights and applying crowding‑distance penalties; step3: Compute selection probabilities based on fitness; step4: Select parents, perform crossover and mutation to create set Q; step5: Merge E and Q, keep the top‑ranking subset for the next generation; step6: Repeat steps 2‑5 until a stopping condition is met.

Application Results

AutoPilot achieves an optimal balance between risk coverage and user disturbance, enabling a shift from a one‑size‑fits‑all approach to individualized control. In O2O offline payment scenarios, it preferentially recommends biometric verification (e.g., facial recognition) for high‑risk cases such as lost phones, thereby protecting account funds.

During the 2017 Tmall Double‑11 promotion, AutoPilot automatically adjusted models and control intensity in response to transaction volume and risk fluctuations, demonstrating fully unmanned strategy adjustment and resilience against black‑market attacks.

Conclusion

AlphaRisk combines human‑intuition AI with machine AI to build a sophisticated risk‑control system. AutoPilot, as its core, delivers scientific decision‑making and “driverless” operation, reducing Alipay’s loss rate from 1/100 000 to less than 1/200 000, far ahead of global competitors.

— END —