Trustworthy AI in the Digital Economy: Practices and Explorations by Ant Group

On March 23, at the Machine Heart AI Technology Conference, Zhou Jun, General Manager of Ant Group's Financial Machine Intelligence Department, delivered the keynote "Trustworthy AI in the Digital Economy: Practice and Exploration," outlining Ant Group's research and deployment of trustworthy AI technologies.

He illustrated the digital economy as a tree, where AI, big data, and cloud computing form the trunk, while privacy and security act as the roots; their tight integration is essential for a flourishing ecosystem, and trustworthy AI is a key capability for risk mitigation and inclusive technology.

Ant Group publicly released its trustworthy AI architecture in June 2020, focusing on privacy protection, explainability, robustness, and fairness, and continues to invest heavily in research and practical applications.

The talk highlighted AI's pervasive role in financial technology, emphasizing the urgent need for AI + privacy and AI + security solutions, and described collaborations with universities to advance these areas.

Graph learning was presented as a core technology, with the Ant Graph Learning (AGL) system comprising three modules—GraphFlat, GraphTrainer, and GraphInfer—designed for scalability, fault‑tolerance, and reuse of existing methods. Large‑scale experiments on a graph with 6.2 billion nodes and over 330 billion edges demonstrated near‑linear speedup using more than 30 000 cores.

Applications of AGL include an anti‑cash‑out detection system that simulates transaction sub‑graphs for link prediction, and supply‑chain mining for SME credit scoring using spatial‑temporal GNN (ST‑GNN) and path‑aware GNN (PaGNN), achieving significant performance gains over baselines such as GBDT and GAT.

To address robustness, Ant Group developed a heterogeneous GNN framework with an attention purifier that mitigates topological adversarial attacks, enhancing AI reliability.

In explainable machine learning, the COCO (Constrained feature perturbation and Counterfactual instances) method was introduced; it generates counterfactuals and importance scores, outperforming SHAP and LIME in accuracy and stability, and is applied to risk‑perception models to reveal decisive factors such as login frequency and transaction anomaly indices.

For privacy‑preserving ML, Ant Group proposed CAESAR, a secure large‑scale sparse logistic regression system built on a hybrid MPC protocol that combines homomorphic encryption and secret sharing, achieving roughly 130× the efficiency of existing SecureML solutions. CAESAR has been deployed in joint risk‑control projects with banks, improving KS metrics by 12‑23% and enabling secure knowledge fusion across institutions.

Adversarial learning efforts include black‑box attack designs and an adversarial training platform that incorporates generated adversarial samples, resulting in smoother decision boundaries, improved robustness, and, in some cases, better accuracy on imbalanced data.

Overall, the presentation underscored the substantial progress Ant Group has made in trustworthy AI across privacy, explainability, robustness, and large‑scale graph learning, while acknowledging that many challenges remain before fully realizing transparent and inclusive AI in the digital economy.