India's DEPA for AI Training: MeitY handbook lets models train on data without moving it

India's IT ministry has published a handbook for a scheme it calls DEPA for AI Training, a framework meant to let hospitals, banks and government agencies make their data available for training AI models without handing over the raw records. The ministry's Secretary, S. Krishnan, unveiled the Handbook on Data Empowerment and Protection Architecture (DEPA) for AI Training on Saturday 31 October 2026 at a roundtable in Bengaluru. The idea, in the words of the government's Principal Scientific Adviser Ajay Kumar Sood, is that "the model travels to the data, rather than the data travelling to the model."

Key points

• MeitY launched the DEPA for AI Training Handbook on 31 October 2026 in Bengaluru, with support from the iSPIRT Foundation (Press Information Bureau release).
• The framework combines contractual permissions, controlled computing environments and privacy-enhancing safeguards, plus data provenance that traces the origin, permissions and use of data through training.
• Under the design described by The Register, data providers share anonymised, encrypted data with exchange platforms that run hardware-based secure environments, and model builders train inside them.
• The Register reports that the ministry's framework document says differential privacy is applied and that all activity in the secure environments is immutably logged for audit.
• It is a framework and handbook, not yet a running national service: The Register says it is unclear how much work remains to make DEPA a reality.

What happened

According to the Press Information Bureau (PIB) release, the Ministry of Electronics and Information Technology (MeitY) launched the handbook at a "high-level roundtable" in Bengaluru. The handbook "sets out a techno-legal framework for enabling trusted AI training while safeguarding data rights and sovereignty". The ministry's argument is that AI needs diverse, high-quality data, while concerns about privacy, confidentiality, intellectual property and institutional control keep organisations from sharing it. DEPA Training is presented as the way around that: purpose-bound data collaboration in which institutions keep control of their data.

The PIB release says the framework has three layers: contractual permissions (what a data holder allows and for what purpose), controlled computing environments (where the training runs), and privacy-enhancing safeguards. It adds data provenance, "enabling the origin, permissions and use of data to be traced throughout the training process".

Krishnan said India's AI ambitions "must be matched by the ability to use data responsibly and securely", and that MeitY is "fully committed" to taking the framework forward. The roundtable heard from HDFC Bank, Infosys co-founder Kris Gopalakrishnan, the AI companies Sarvam and Emergent AI, and iSPIRT co-founder Sharad Sharma, who called it "easily a more significant moment" than any since the announcement of the Unified Payments Interface (UPI) in February 2016. That is Sharma's view, offered at the launch, not an independent assessment.

How it works

The Register, which read the ministry's framework document (a PDF), describes the roles like this:

• Training Data Providers (TDPs): organisations that hold data, for example hospitals, banks and government agencies. They share anonymised and encrypted data with a Data Exchange Platform.
• Data Exchange Platforms (DEPs): the intermediaries. They operate what MeitY calls "hardware-based Secure Data Environments", in which data "remains encrypted and isolated even during processing", and they enforce the rules each provider sets for how its data may be used.
• Training Data Consumers (TDCs): the AI developers. They access the data stored at the exchanges to train their models, inside those environments.

Two safeguards are named in the framework, as quoted by The Register. First, "Differential privacy techniques are systematically applied to prevent the model from learning or revealing individual specific information". Differential privacy is a standard technique that limits how much any one person's record can influence a trained model. Second, "All activities inside the secure environments are immutably logged for audit purposes".

The phrase "hardware-based secure environment" points to what the industry calls confidential computing: processors that keep data encrypted in memory and let only approved code touch it. The sources we read do not name specific hardware, vendors or software, so we do not know which technology India expects exchanges to use.

The practical effect of "model to data" is that the data holder never exports a dataset. A developer sends a training job into the environment and, under rules the holder sets, receives a trained model back. The holder can decline purposes it does not approve, and the logs let it check what ran.

Background

India has spent a decade building what it calls digital public infrastructure (DPI): government-backed open systems such as UPI that private firms build on. The Register notes the framework mentions DPI repeatedly, and that the stated aim is to keep Big Tech companies from dominating the local market. HDFC Bank's Anjani Rathor told the roundtable that the bank's experience with "DEPA Inference" showed the promise of secure data collaboration and that DEPA Training builds on it.

The new handbook arrives amid a global argument over whether AI developers may use data without paying for it. The Register frames the scheme as India's attempt to "regulate and secure the flow of information", with the motive of creating models trained on local data, because, it reports, India thinks offshore model-makers "won't be motivated" to address the country's issues or leave data owners in control. That motive is the Register's reading, not a line from the ministry.

What it means

For data holders. The offer is control: a hospital or bank could contribute to model training without moving its records. The framework's examples are health data for diagnostics and drug discovery and financial data that could extend formal credit to 400 million underserved individuals and small businesses. In education, it points to personalised learning for 250 million students in a $29 billion edtech sector. Those figures are the ministry's own projections from its framework, as reported by The Register, and we have not been able to verify them independently.

For AI developers. The trade-off is access versus freedom. Training inside someone else's locked environment, under their rules, is harder than downloading a corpus. Debugging, evaluation and iteration happen with limited visibility into the data. No company has announced that it is training through DEPA.

For privacy. Moving the model rather than the data lowers the risk of a copied dataset leaking. It does not remove the risk that a model memorises something. That is why the framework leans on differential privacy, whose strength depends on parameters the handbook, as reported, does not detail in the coverage we read.

What we don't know yet

• What is live. Sources describe a launched handbook and framework. We found no report of an operating exchange, a pilot dataset or a start date.
• Who may run an exchange and who certifies the secure environments.
• Legal footing. The coverage describes "contractual permissions" but does not say how the scheme interacts with India's data protection law.
• Cross-border use. Sood said DEPA enables "cross-border data collaboration"; the sources do not explain the rules for foreign developers.
• Privacy parameters and hardware. Not stated in the coverage.

What to watch next

• Announcements of the first Data Exchange Platforms and pilot sectors, most likely health and finance.
• Whether Indian model builders such as Sarvam say they will train through the scheme.
• The full handbook text and any technical specifications for the secure environments.
• Whether other governments adopt the "model to data" approach, as Sood suggested it could for global science collaboration.

Sources

• MeitY Launches DEPA for AI Training Handbook, Press Information Bureau (31 October 2026)
• India floats AI plan under which "the model travels to the data", The Register (2 November 2026)
• Govt launches handbook on DEPA for AI training, ETV Bharat
• MeitY launches DEPA handbook, Business Standard (PTI)
• MeitY launches DEPA handbook, The Tribune (ANI)