Taiwan AI Labs Federated Learning Platform

TAIMedimg solution of collaborative research

Federated learning (FL) is a machine learning method that enables multiple parties to collaboratively train a model orchestrated by a trustable central server while keeping data locally. Taiwan AI Labs built up the first open-source federated learning framework (FL framework) in Taiwan. AILabs FL framework is widely deployed and used across medical centers and regional hospitals in Taiwan. It obtains huge reputation due to its privacy protection mechanism, cross-institutional collaboration, including model training (Federated Learning, FL), validation (Federated Validation, FV®).

Key Features

Holistic data governance tools including curation, quality control, preprocessing and labeling. Providing unified normalization approaches to standardize data excellence and usage.

To guide the responsible AI development, we ensure the framework safety, a full spectrum of data governance, including data ETL process, de-ID, quality control, preprocessing and usage to protect data privacy, to offer accountability to guarantee human rights.

Data Extract, Transform, Load

Data Extract, Transform, Load


Ground-Truth Consensus


Quality Control

Diverse Data Format Annotation

Privacy protection first – complying with GDPR by data deidentification, anonymization, pseudonymization with user consent.

Complying the human rights to be informed, restrict, object and erasure, the platform exams imported data to ensure data de-identification, and removes individual data upon requests.

Simultaneous multiple-site, multiple-project training processes, with managing dashboard to create, join projects, to upload AI model but keep datasets locally, and to setup with flexible schedule and fault-tolerant network. Advanced aggregation algorithms for better performance and best local model output.

Federated learning keeps data in private while training AI models, to maximize data availability and diversity. This fundamentally enhances the data-insufficient and data-bias dilemma in democracy, which provides a solution for fairness and non-discrimination of AI model.

Cross-site, cross-vendor federated validation mechanism provides an easy solution to validate AI model to avoid data bias and discrimination.

FV offers a mechanism to evaluate AI model performance across institutes without disclosing data. With wide range of diverse data validation, the robustness and reliability could be assured and convinced by public. Accompanying CRO reports during FV progress help speed up FDA progress.

Taiwan AI Federated Learning Alliance (TAIFA, https://taifa.org/) assembles more than 90 members from medical and other industries, government, and academy. They form Federated Learning Alliances in areas such as medical, financial, transportation, manufacturing, etc., to advocate and implement Federated Learning in order to solve data silo issues.

TAIFA was initiated by all key industrial ministries of Taiwan Central Government. More than 90 industry/government/academic/hospital leaders are involved together to advocate the Next Generation Trustworthy Federated Learning Paradigm with Data Governance.

Supported by central and local governments, TAIFA promotes Federated Learning in various fields such as medical/finance/smart city/cultural/business/ manufacturing. Through alliance government-industry-university-institute network, to speed up cross-domain innovation ecosystem formation, to accelerate matching and policy support, to connect allies efficiently, and to gathers the power of members to obtains most possible large amount of diverse data, to conduct cross-organization AI model training and validation. Meanwhile to realize software and hardware integration of AI landing solutions through alliance manufacturers.

Cooperate with the Taiwan Clinical Trial Consortium (TCTC) for field validation while customizable tools are provided for CRF/CSR reports during CDE/TFDA/FDA SaMD clinical evaluation progress.

In SaMD process, we provide the high effective PI matching and site selection for our customers. We are cooperating with TCTC (https://www.taiwanclinicaltrials.tw/tctc/member, Taiwan Clinical Trial Consortium), which has the biggest country level PI pool in Taiwan. TCTC is subordinate to TFDA.

In TCTC, there are more than 300 clinical trial experienced doctors under 20 medical centers divided by disease categories, including but not limited to oncology, infection and gastroenterology. Medical centers in TCTC use FV systems to follow the TFDA central IRB process, which shortens the validation progress effectively.

From data collection, annotations, consensus, to training and validation, each process step is signed off with a FL certificate. The process is auditable and traceable, in order to prevent malicious attack and poisoning, and to measure individuals and institutions’ contributions to AI models.

Immutable Signoff records the whole progress of AI model development, from dataset selection, annotation, model choice, and training result. The unchangeable digital certificate provides the full history of each AI model development, offers transparency and traceability for each AI model.

Signoff at All Steps


Signoff Management

Signoff During FL:FV Progress

AI Production Certification for Traceability

ISO/IEC 27000, ISO/IEC 29000 and CNS 29100 privacy compliance. OWASP (open web security project) SAST (static application security testing) and DAST (dynamic application security testing) ensure.

Taiwan AILabs FL Platform ensures code security and environment robustness by applying industry highest standard security review, which meets OWASP (Open Web Application Security Project) Top 10, ISO 27001 / 270101 / 29100 requirements rigorously.

ISO, OWASP Compliance


ISO Privacy OWASP Top Ten

Security Compliance

A complete set of API connectors to link hospital HIS/PACS/RIS/FHIR/OHDSI/CDISC/MIMIC-IV/… systems. Effortless deployment to streamline data flow.

Taiwan AI Labs provides a complete set of API calls to communicate with the Federated Learning Platform.

The data is transferred streaming at real-time.