On December 9, the results of the 2022 iDASH International Privacy Computing Competition were officially announced. Tencent Angel PowerFL United won the championship of homomorphic encryption track with its best model effect and fast inference speed.

On December 9, the results of the 2022 iDASH International Privacy Computing Competition were officially announced. Tencent Angel PowerFL United won the championship of homomorphic encryption track with its best model effect and fast inference speed. From the first time participating in the competition two years ago to the present, Tencent has won the 2020 Trusted Computing Track Championship and the 2021 Federal Learning Track Championship, winning the "three consecutive championships" of the iDASH Privacy Computing Competition, which is the first time in China!

At the same time, Tencent Angel PowerFL United also achieved second and third good results respectively in the multi-party security computing (MPC) track and trusted computing (SGX) track.

Over the years, homomorphic encryption track has been the most popular and fiercely contested track in iDASH competitions. This year's iDASH homomorphic encryption track problem is security model reasoning, requiring participating teams to train machine learning models and predict phenotypes through genotype data. There are five prediction tasks, including three regression tasks and two classification tasks. The main challenge is that both model parameters and test data must be encrypted and protected, and model inference needs to be completed under the ciphertext.

For five prediction tasks, Tencent Angel PowerFL Wing trained three linear regression models and two logistic regression models on the data set disclosed by iDASH, and obtained model effect indicators that were close to full scores. When performing security model inference, the parameters of the five linear models and the data matrix to be tested are encrypted using the CKKS homomorphic encryption algorithm, model inference is completed under the ciphertext, and the optimal model inference speed under a single thread is obtained by optimizing the ciphertext matrix and ciphertext vector multiplication.

It is understood that this year the iDASH homomorphic encryption track has attracted more than 30 top teams from all over the world to sign up for the competition, including participating teams from Yale University , EPFL, Zhejiang University , Intel, Ant Group, ByteDance and other institutions. Among them, 23 participating teams submitted technical solutions and implementation codes, and obtained effective competition results. Tencent Angel PowerFL proposes multiple innovative solutions and ultimately obtains the highest comprehensive score in terms of security model inference results and inference speed.

This year's iDASH multi-party safety computing track question is safety record association, requiring participating teams to be safely related to the records of the same patient in two databases. The main challenge is that the patient record information may be missing and errors, and no patient information cannot be disclosed. Tencent Angel PowerFL Wing innovatively proposed a machine learning-based solution, trained a logistic regression model on the data disclosed by iDASH, and implemented full anonymized model inference based on circuit privacy interception (Circuit-PSI) and obfuscating switching network (Oblivious Switching Network), achieving the highest accuracy rate. The solutions proposed by Tencent are widely applicable and can be applied in financial, government affairs and other scenarios.

In addition, this year's iDASH trusted computing track problem is to safely cluster cells based on the similarity of cell gene fragment distribution. The main challenge is to implement distributed secure clustering solutions in multiple Intel SGX Enclave environments. The Message Queue-based distributed solution proposed by Tencent Angel PowerFL Wing is the only solution among all award-winning teams that can run on multiple machines. It is a true distributed solution with good scalability and supports massive data computing. It can be applied to large-scale distributed trusted computing based on SGX in production environments. The

iDASH competition is currently the most authoritative international competition in the field of privacy computing. It has been held for nine sessions so far. It is hosted by National Institute of Health (NIH). It focuses on privacy computing and machine learning issues for privacy protection. It has become the highest-standard international competition in the field of privacy protection and security sharing of genomic data in the world.

This year's Tencent Angel PowerFL team brought together technical experts from Tencent Big Data, Tencent Security, Tencent Billing, Tencent Cloud , Tencent Advertising AI, Huazhong University of Science and Technology , cryptography , privacy computing, big data and machine learning technology experts from Tencent Big Data, Tencent Security, Tencent Billing, Tencent Cloud , Tencent Advertising AI, Huazhong University of Science and Technology , cryptography , , .Tencent Angel PowerFL privacy computing team is the earliest team in China to carry out research and application of privacy computing and federated learning technology. It has rich R&D and application experience in big data, distributed computing, distributed machine learning, distributed message middleware, multi-party security computing, applied cryptography and other fields. It has published nearly 10 privacy computing research papers and submitted more than 60 privacy computing technology invention patent applications. There are multiple commercial privacy computing and federated learning platform products that have been opened to the public through Tencent Cloud.