How Tencent’s Angel PowerFL Team Dominated iDASH with Homomorphic Encryption

Tencent announced a major breakthrough in privacy computing as its Angel PowerFL team won the iDASH homomorphic encryption track, while also achieving second place in the multi‑party computation (MPC) track and third place in the trusted execution environment (SGX) track.

The iDASH International Privacy Computing Competition, organized by the U.S. National Institutes of Health, is the most authoritative global contest for privacy‑preserving machine learning and secure data sharing. This year’s homomorphic encryption track focused on secure model inference, requiring participants to train models that predict phenotypes from genotype data under full encryption.

Tencent’s team trained three linear regression and two logistic regression models on the public iDASH dataset, attaining near‑perfect performance metrics. For inference, they encrypted model parameters and test data using the CKKS homomorphic encryption scheme and performed ciphertext‑level computation, optimizing matrix‑vector multiplication to achieve the fastest single‑thread inference speed.

In the MPC track, the challenge was secure record linkage across two databases without exposing any patient information. Tencent’s solution combined a machine‑learning approach with Circuit‑PSI and an Oblivious Switching Network to achieve fully hidden model inference and the highest accuracy, applicable to finance and government domains.

The SGX track required secure clustering of cells based on gene fragment similarity across multiple Intel SGX enclaves. Tencent’s unique distributed solution built on a message‑queue architecture was the only multi‑machine implementation among the winners, offering strong scalability for large‑scale trusted computing workloads.

The Angel PowerFL team brings together experts from Tencent’s big data, security, cloud, advertising AI divisions and researchers from Huazhong University of Science and Technology, covering cryptography, privacy computing, federated learning, distributed systems, and applied cryptography. The team has published nearly ten papers, filed over 60 patents, and released commercial privacy‑computing platforms on Tencent Cloud.

—END—