India: the model travels to the data, not the reverse

What is it?

On October 31 in Bengaluru, MeitY presented the DEPA for AI Training handbook, developed with support from the iSPIRT Foundation. The document is still marked "Draft for Discussion". Principal Scientific Adviser Ajay Kumar Sood summed it up: "the model travels to the data, rather than the data travelling to the model".

How would it work?

The two main roles are the Training Data Provider (TDP), such as a hospital, bank or public agency, and the Training Data Consumer (TDC), such as a startup, research group or enterprise. Between them sit Data Exchange Platforms, for example the government's AI Kosh. The process has five steps:

• Registration: the TDP anonymizes and encrypts the data and tags it with metadata and cryptographic markers for end-to-end traceability.
• Request and contract: the TDC requests access for a specific use case. Machine-readable electronic contracts are validated automatically.
• Training in a secure environment: inside hardware-based Secure Data Environments, data stays encrypted and isolated even during processing. Training uses differential privacy, and every action is immutably logged.
• Validation: bias and fairness audits, security and robustness testing, and extra testing for high-risk applications.
• Deployment and continuous monitoring, with traceability from model output back to the data and contracts.

What India wants from it

The goal is models trained on local data for local problems. The ministry argues offshore developers have little incentive to solve India's issues, while data owners keep control under DEPA. The document cites health, finance and education, claiming secure sharing of financial data could extend formal credit to 400 million underserved people and small businesses. India also frames DEPA as a contribution to global "science diplomacy", following the model of its digital public infrastructure (DPI).

The open questions

For now this is a concept. It's unclear how much work remains before the platforms, the contract management system and the secure environments actually run. Strong privacy techniques can also cost model quality. If it works, though, it's one of the most concrete government answers to how AI can be trained on sensitive institutional data without handing that data over.