Serpro Tests Split Learning So Two Institutions Can Train One AI Model Without Pooling Their Data
Serpro, the federal government’s data processing company, has shown how two institutions can train the same AI model while each keeps its data in its own environment, Inforchannel reported on November 21. At the 2nd Serpro AI Week, the state company split a public dataset of card transactions to simulate a bank and a card processor, and only the processor was told which purchases had been classified as fraud. Both sides then trained a single model. Serpro is studying whether to apply the approach, part of the family of techniques known as federated learning, in a pilot with Receita Federal, the federal tax authority.
In the simulation each participant ran part of the neural network, a setup called split learning. Serpro specialist Alexandre Henrique Soares explained that only the embeddings, which are intermediate representations, and their gradients travel between the participants, and those calculations adjust the model at each training round. The raw records stay where they are. Soares also warned about evaluation: because fraud made up a small share of the transactions, a model can be right most of the time and still miss exactly the purchases it should flag.
What it means in practice
The accuracy point was taken up by Camila Falchetto Quadros, general coordinator of AI governance at the Digital Government Secretariat of the Ministry of Management and Innovation in Public Services (MGI). Using a hypothetical benefits system that is right 95% of the time, she asked who falls in the other 5%, whether they form a group and whether they are being discriminated against. To explain to a citizen why a benefit was refused, she said, an agency must be able to reconstruct the path to the decision, and civil servants must be able to review results, correct errors and stop the system when needed. In the 2025 self-diagnosis, 120 federal bodies said they used AI or had started adopting it, but only 15 carried out risk assessments.
That gap is what the ethical impact self-assessment now online for more than 250 federal bodies is meant to close, with 11 questions for low-risk systems and 44 for high-risk ones, and Serpro helped calibrate it. Federal agencies are also preparing more ambitious uses, such as transactional AI for Contrata+Brasil. What to watch is whether the Receita pilot goes ahead and what data it would cover.