MeitY unveils a DEPA handbook for AI training, proposing sealed-room access to sensitive Indian datasets
MeitY's draft DEPA for AI Training framework, built with iSPIRT, would let developers train models on bank, hospital and government data inside secure environments with mandatory differential privacy, with controlled access to sensitive data.
India's IT ministry has put forward a blueprint for one of the hardest problems in building home-grown AI: how to train models on sensitive data held by banks, hospitals and government bodies while protecting institutional control over sensitive data. In its 31 October 2026 launch report, at a roundtable in Bengaluru, Ministry of Electronics and Information Technology (MeitY) Secretary S Krishnan unveiled the Handbook on Data Empowerment and Protection Architecture (DEPA) for AI Training, prepared with support from the volunteer-run iSPIRT Foundation.
Why India needs this
Every AI model is shaped by the data it learns from. Indian developers building models for Indian languages, health systems and financial services face a basic shortage. The most valuable datasets — patient records, loan histories, insurance claims — sit inside institutions that have little reason to share them and strong reasons not to.
The result is that Indian AI start-ups often depend on public web data or foreign datasets that reflect Indian realities poorly. DEPA for AI Training is the government's attempt to unlock domestic data while keeping legal and privacy risks manageable.
What the framework proposes
The handbook describes a "techno-legal" framework: technology and contracts working together so that rules are enforced by the system itself. It rests on four pillars:
- A Data Exchange Platform that connects data providers with AI developers ("data consumers").
- Machine-readable electronic contracts that state exactly what a developer may do with a dataset.
- A Secure Data Environment — described in coverage as a "sealed room" — where training runs, and which executes only what the authorised electronic contract permits.
- Differential privacy, a mathematical technique that adds calibrated noise to bound the influence of individual records, with privacy and utility depending on implementation and parameter choices.
The handbook describes two ways for organisations to work together. In the first, data providers anonymise their own data, and consumers can either download it or work on it inside a secure environment. In the second, the Data Exchange Platform takes care of anonymisation and encryption, and developers can only reach the data inside the secure environment. In both models, the raw data stays out of developers' hands.
From consent to collaboration
The AI Training version adds a separate layer for a different job: making large volumes of de-identified data available to build models. It combines contracts, controlled computing and privacy safeguards; the handbook does not itself waive applicable consent requirements or other legal duties.
Who was in the room
The launch drew a senior group. Principal Scientific Adviser Prof Ajay Kumar Sood, Infosys co-founder Kris Gopalakrishnan and iSPIRT co-founder Sharad Sharma took part, along with government, academic and industry representatives. That mix — a large private bank, a sovereign-model developer and an AI start-up — reflects the two sides the framework is trying to bring together.
What is still open
The handbook is a draft and is not legally binding. The document is marked “Draft for Discussion”; its proposed technical architecture is not evidence of an operational nationwide service.
Several practical questions will decide whether it works:
- Incentives. A secure pipeline lowers the risk of sharing, but banks and hospitals still need a reason to take part. Revenue-sharing, recognition or regulatory credit may have to follow.
- Compute and cost. Secure environments with differential privacy are more expensive to run than ordinary training clusters. Who pays for that, and whether it will draw on IndiaAI Mission compute, is not yet clear.
- Model quality. Differential privacy involves a trade-off: stronger privacy usually means somewhat less accurate models. The settings chosen in the pilots will matter.
- Fit with the DPDP regime. Organisations will want certainty that following DEPA protects them from liability under the Digital Personal Data Protection framework.
Why it matters beyond India
India has already exported the logic of digital public infrastructure through UPI, Aadhaar-based identity and Account Aggregators. If DEPA for AI Training works in pilots, it could become a reference model for other countries that want to build AI on local, sensitive data without handing it to a few global platforms. For Indian AI developers, the immediate prize is simpler: lawful, privacy-protected access to the kind of data that makes models truly useful for Indian users.
Sources
- https://www.pib.gov.in/PressReleasePage.aspx?PRID=2330756®=48&lang=2
- https://www.tribuneindia.com/news/business/meity-launches-depa-handbook-to-enable-secure-data-sharing-for-ai-training/
- https://www.business-standard.com/amp/technology/artificial-intelligence/india-proposes-framework-to-unlock-data-for-ai-training-protect-privacy-126103001392_1.html
- https://yourstory.com/2026/10/meity-unveils-depa-handbook-to-enable-trusted-data-sharing-for-ai-training
- https://cio.eletsonline.com/artificial-intelligence/meity-depa-ai-training-framework-sealed-rooms-sensitive-data-ispirt/76548/
- https://www.etvbharat.com/en/bharat/govt-launches-handbook-on-depa-for-ai-training-enn26103101486
- https://www.meity.gov.in/static/uploads/2026/10/02d8224e4cd6c8c29be5a5198e795158.pdf