Data sovereignty also requires security, quality controls and verifiable measures of each participant’s contribution
In This Series
Part 1 — Data as National Infrastructure
Part 2 — How Data Becomes an Industrial Asset: Yang Jae-soo’s Framework for Verification
Part 3 — DataStreams Takes Trusted, AI-Ready Data Ecosystems Global
Part 4 — FLock.io CEO Jiahao Sun on Keeping Data Local While AI Learns
Part 5 — Causal Dynamics Lab CEO Hasibul Haque: “AI Agents Need Operational Context and Accountability”
Part 6 — KIT Analysis: Korea Must Prove Its Data Competitiveness Through Verifiable Industrial Outcomes
Centralizing Sensitive Data Is Becoming a Constraint on AI Expansion
Data held by governments and companies is a critical source of AI competitiveness, but transferring personal information, trade secrets and national security data to centralized servers creates significant risks. Healthcare providers, financial institutions and public agencies need ways to develop shared AI models while retaining control over their data.
Federated learning sends an AI model to the organizations or devices where data resides instead of bringing the data together in a central location. Each participant keeps its raw data within its own environment and shares only model updates. This approach is gaining attention in healthcare, finance, manufacturing and the public sector because it allows organizations to develop a common model without directly exchanging sensitive information.
Jiahao Sun, founder and CEO of FLock.io, told Korea IT Times in a written interview that the scarcity and concentration of high-quality data, along with the legal risks surrounding its use, are constraining AI expansion. He said federated learning can enable institutions and countries to collaborate while retaining control over their data. Combining it with blockchain-based records and incentive systems could also help assess the value of data contributions and hold participants accountable, he added.
Keeping raw data within an organization does not, however, provide complete privacy protection or data sovereignty. Secure aggregation, differential privacy and encryption may be needed to reduce the risk of information leaking through model updates. Organizations must also determine how to evaluate data of differing quality and allocate the performance gains and economic value produced by a jointly trained model.
The following is a Q&A with Jiahao Sun, founder and CEO of FLock.io.
Q1. Why is federated learning becoming increasingly important to national and corporate AI strategies?
The importance of federated learning, or FL, in today’s AI landscape, both national and corporate, cannot be overstated. Traditional machine learning models have largely exhausted publicly available datasets, reaching a plateau in performance improvements. To push the boundaries further, there is a pressing need to incorporate local and proprietary data, which holds rich, untapped potential for training more accurate and robust models.
As compute costs stabilize, the true constraint in scaling AI has shifted from GPUs to the scarcity, centralization and legal risks surrounding high-quality data. FL transforms how we train models, allowing collaborative training while protecting data privacy and sovereignty.
For national strategies, two main considerations relate to data security and sovereignty.
First, centralized AI presents a problem because data is sent to a central cloud server belonging to corporations that are often based overseas. This compromises privacy and national security by surrendering datasets such as citizens’ demographic data, military intelligence and healthcare records. With FL, data can remain within the country’s jurisdiction and on state-owned hardware. This allows collaboration with allied nations, organizations and academic institutions by sharing only model updates and never the raw data.
Second, nations are pursuing sovereign AI so they can retain control over their own AI. There is a reason we use the phrase “Not your models, not your AI” on FLock.io’s homepage. When the U.S. government temporarily suspended access to two Anthropic models for non-U.S. citizens in June, it served as a wake-up call to governments that sovereign AI must form a key part of their national AI strategies. AI is no longer simply a tool. It is becoming critical infrastructure, as essential to the functioning of society as the internet, oil or electricity.
Corporations are embracing FL because of the regulatory, legal and operational risks presented by centralized AI. Leading law firms, for example, are moving toward owning custom-built, proprietary AI infrastructure instead of merely renting software. This is necessary because such firms are bound by strict client-confidentiality requirements. FL enables them to build this infrastructure without data ever leaving the firm.
Q2. How can organizations train AI across distributed datasets without exposing sensitive raw data?
Organizations can overcome the data bottleneck by reversing traditional machine learning workflows. Instead of bringing the data to the model, FL brings the model to the data.
A global model is dispatched to local edge nodes or internal enterprise servers where the raw data resides. Each participating node trains the model locally using its isolated dataset. The underlying raw data never crosses that boundary. Instead, the local node calculates abstract mathematical parameter updates, such as gradients and weight updates, that reflect what the model has learned.
To prevent these parameter updates from being reverse-engineered to reconstruct raw source data, FL incorporates privacy-preserving techniques. These include differential privacy, homomorphic encryption and secure aggregation.
Q3. How should data ownership, privacy and the economic value of data contributors be protected in this model?
Standard federated learning has a key vulnerability: it relies on a single centralized coordinator that holds absolute power over the global model and determines how value is distributed. At FLock.io, we address this by combining FL with blockchain to create a decentralized economy through our AI Arena, FL Alliance and FOMO platforms.
Blockchain-based rewards encourage the contribution of high-quality data. Contributors can be rewarded or penalized according to the quantifiable effect their data has on improving or harming a model. This vision of a fair and incentive-compatible ecosystem is implemented through a reward-allocation system designed to promote accountability and generate active demand for the token.
The value of the FLOCK token corresponds to the value of the resources consumed during these processes. If a participant is found to have acted against the system’s rules or attempted to undermine it through malicious activity, that participant is subject to “slashing.” The forfeited tokens are distributed to honest participants as rewards.
Federated learning protects data ownership and privacy by keeping raw data local rather than sending it to a central server. Multiple edge devices collaboratively train a single global model while sharing only model updates. Ownership rights can be verified through on-chain governance because they are recorded in smart contracts.
FOMO, or FLock Open Model Offering, connects trained models with their deployment and use in real-world services across the FLock ecosystem.
Q4. What lessons could Korean companies and public institutions draw from international sovereign AI projects?
South Korea has strong digital infrastructure, major industrial groups and distinctive domestic datasets. As the Ministry of Science and ICT pursues an independent AI foundation model to position the country as an AI powerhouse and reduce its reliance on global models, it should incorporate lessons from projects elsewhere.
Better models require better datasets. International cases show that sovereign capabilities built on decentralized networks can allow regional medical centers, financial institutions and public utilities to co-train national models while retaining sensitive records within their own environments. FLock.io is working with Moorfields Eye Hospital in the UK on eye-disease detection and with University College London Hospitals on glucose management for patients with diabetes. The latter initiative was also highlighted by the World Economic Forum’s MINDS program. With decentralized infrastructure, Korean institutions can comply with strict data regulations while expanding AI training and services.
In the Dominican Republic, FLock.io is working with the United Nations Development Programme to advance public-sector digital innovation. The collaboration builds on a pilot conducted through the second cohort of the SDG Blockchain Accelerator, in which FLock.io participated as a strategic partner. Korea can draw lessons from efforts to develop data-driven public services while promoting international knowledge exchange, institutional dialogue and shared learning.
Korean has less high-quality data available for AI training than English, Mandarin or Spanish, making it difficult to secure sufficient material to improve model performance. Its vocabulary and grammatical characteristics also create challenges for developing accurate models.
Korea can draw lessons from the sovereign AI initiative in Sarawak, Malaysia’s largest state. FLock.io is the project’s key technical partner and says it developed the world’s first use of federated learning to train and fine-tune a language model for Indigenous languages. The same approach could be applied to Korean and other languages with fewer data resources than the world’s most widely used languages.
What It Means
Jiahao Sun's responses point to the potential for governments and businesses to expand AI training while retaining control over their data. Federated learning allows hospitals and companies to manage where raw data is stored and who can access it, while contributing to the development of models trained across multiple institutions. This could bring data previously constrained by privacy and trade-secret concerns into AI development across healthcare, finance, manufacturing and public services.
Federated learning needs additional safeguards to strengthen trust in the exchange of model updates. Secure aggregation, differential privacy and homomorphic encryption can reduce the risk of exposing source data during collaborative training. Sector-specific testing is also needed to assess their effects on computing costs and model accuracy, enabling organizations to manage data protection and AI performance together.
FLock.io rewards participants that contribute high-quality data with tokens and applies “slashing,” which forfeits tokens deposited by participants that violate rules or attempt to compromise a model. This operating model seeks to translate the value of data contributions into economic rewards while holding participants accountable. Transparent rules for rewards, decision-making authority and appeals could support a more sustainable collaborative-training ecosystem.
The Moorfields Eye Hospital, UCL Hospitals, UNDP–Dominican Republic and Sarawak projects in Malaysia indicate how federated learning could extend into healthcare, public services and AI development for lower-resource languages. The publicly available information does not yet provide enough detail to independently assess the number of participating institutions, data volumes, model performance or operational outcomes. Further disclosure of project scope and performance metrics would provide a stronger basis for evaluating the technology’s industrial impact.
Korea could apply federated learning to national data infrastructure to protect sensitive information while expanding collaboration across institutions and industries. This will require records identifying who requested each training task, what data was used and how the model changed, along with systems for evaluating data quality and each participant’s contribution. The ability to govern storage, access, training processes, economic value and accountability can turn data sovereignty into industrial competitiveness.
The next installment features Hasibul Haque, CEO of Causal Dynamics Lab, on the conditions AI agents need to understand enterprise production environments and act responsibly. It examines how a Production World Model connects code and systems with security policies, ownership and incident history. It also explores how a Decision Trace records the evidence and permissions behind an AI agent’s decision, who approved it and what happened after deployment.
