Ant Group's Edge‑Cloud Collaborative Risk Control: Architecture, Algorithms, and Privacy Protection

The development of IoT, cloud services, big data, AI, and 5G has created a demand for greener, more efficient, and privacy‑preserving solutions, and edge‑cloud collaborative technology meets these needs.

At the AIOT2023 Intelligent Edge‑Cloud Computing Forum, Ant Group’s full edge‑cloud collaborative technology was disclosed by Fu Xinyi of the Ant Group Security Machine Intelligence Department.

Key challenges include the difficulty of precise risk identification due to increasingly sophisticated black‑market behavior and the need for real‑time payment risk judgment without degrading user experience.

New business models such as the mini‑program ecosystem generate decentralized data, requiring responsible detection of content violations and illegal operations on merchants’ private servers.

Traditional cloud‑centered risk control incurs high cost, load, and privacy concerns; edge intelligence offers distributed computing, keeps sensitive data on the device, and reduces latency, providing three main advantages.

Ant Group’s evolution progressed through three stages: pure cloud, edge‑cloud (partial inference on the client), and edge‑cloud‑edge (adding CDN/edge servers) to balance cost, privacy, and performance.

In payment risk control, billions of real‑time features and hundreds of rules create massive load; Ant introduced a “global trust” concept to quickly pass low‑risk transactions, reducing downstream computation.

Three models are deployed on device, edge, and cloud; a real‑time scheduler estimates cost and value to decide where inference runs, with results aggregated for final judgment.

For private content risk in the digital ecosystem, Ant uses on‑device element‑recognition models to detect fraud while keeping user data local.

Core algorithms include “thousands of people, thousands of models” with personalized parameters and structures, meta‑learning a base model then fine‑tuning per group, and model compression that selects device‑specific sub‑models based on performance profiling.

Privacy protection is achieved through data minimization via federated/split learning, representation encryption, and privacy measurement using mutual information; deployment privacy uses differential‑privacy training with gradient clipping and noise, lowering membership‑inference success rates.

Edge‑cloud collaboration enables greener, responsible AI handling unprecedented data and compute demands, and suggests layered regulation of data collection, storage, and usage.

Ant’s edge‑cloud risk‑control products will be showcased at the Bund Conference from September 7‑9, inviting readers to visit the “Trusted Edge Devices” exhibition.