India’s Ministry of Electronics and Information Technology (MeitY) has published DEPA for AI Training, a draft framework for pooling bank, hospital and government data to train AI models inside secure enclaves where raw records stay hidden. The document, posted on MeitY’s website in late October 2026 and reported by AI News on November 2, was prepared with the iSPIRT Foundation and is marked as a draft for discussion. Its central promise is that developers learn from the data without ever seeing it. One of its two deployment options lets developers download the data anyway.
DEPA for AI Training Rests on Four Pillars and a Five-Step Workflow
The DEPA for AI Training draft builds on four components: a Data Exchange Platform, with India’s AI Kosh named as the candidate; machine-readable electronic contracts that encode the terms of each data use; Secure Data Environments, which are hardware enclaves where computation happens; and differential privacy, which adds calibrated statistical noise under audited privacy budgets. Its workflow runs in five steps. A Training Data Provider registers, anonymises, encrypts and watermarks a dataset. A Training Data Consumer requests it, and the contract system checks the request against the provider’s policies. Training then runs inside the enclave with differential privacy applied and every action logged. The trained model goes through independent validation for bias, robustness, safety and explainability. Deployment follows, with continuous monitoring and traceability back to the source data and contracts.
The DEPA Draft Says Individual Consent Is Impractical for Pooled Training
India already runs a consent system for personal data in finance, the Account Aggregator network, which the draft calls DEPA Inference. It lets a person approve each transfer of their bank or tax records for a specific purpose. The draft argues that per-transaction consent cannot work when several institutions want to pool millions of records to train a model, and that simple anonymisation fails because of re-identification risk and lost utility. Its replacement is institutional: de-identification before sharing, with pseudonymised keys held by the originating agency, plus contracts, enclaves and privacy budgets. Intermediaries the draft calls Data Collabs may be responsible for obtaining primary consent where it applies.
Option 1 Lets Developers Download Provider-Anonymised Data
The draft offers two ways to run the system. Under Option 1, each data provider anonymises its own data, and developers can download it or use the exchange’s secure environment for training. Under Option 2, the exchange anonymises and encrypts the raw data itself, and developers reach it only inside the enclave. Option 2 matches the draft’s own description of a strict no-exposure policy for raw data. Option 1 puts the privacy guarantee back on the provider’s anonymisation, the same technique the draft says breaks down at scale. The document leaves both options open and does not say which sectors or data classes would use which.
A Government Authority Would Certify Models Before Deployment
The draft gives a competent government authority the job of registering enclave providers, verifying trained models as fit for use, and suspending or revoking non-compliant activity. Its roadmap starts with pilot exchanges in finance and healthcare, regulatory sandboxes for testing models on DEPA datasets, and incentive schemes such as royalties or access fees for data providers. As evidence that enclave training works, it cites a Penn Medicine and Intel federation of 71 institutions whose brain tumor model reached over 99% of the accuracy of a centrally trained equivalent. The draft adds that the greatest barriers in that project were non-technical.
Why This Matters for Engineering Leadership
The choice between the two options decides who carries the privacy risk, and that choice is still open. For a bank or hospital engineering team that might become a Training Data Provider, Option 1 makes anonymisation your job, and the draft itself says simple anonymisation fails on re-identification risk. Plan for it as a product with re-identification testing, ownership and an audit trail, or argue in the consultation for Option 2, where the exchange carries that work and data never leaves the enclave. For teams that would consume DEPA data, design training pipelines that run inside a remote enclave with no local copies, treat the privacy budget as a finite resource you schedule like compute, and expect a certification gate before any model ships. The document invites recommendations from individuals and institutions, so this is the window to say which option your organisation can actually operate safely. Teams outside India can borrow the same test for any data-sharing deal: find the clause that allows a download, because that clause defines the real privacy guarantee.
FAQ
What is DEPA for AI Training?
It is a draft framework from India’s Ministry of Electronics and Information Technology, prepared with the iSPIRT Foundation, for sharing aggregate datasets to train AI models without exposing raw records.
Is DEPA for AI Training in force?
No. The document is marked as a draft for discussion and invites recommendations from individuals and institutions. Its roadmap starts with pilots in finance and healthcare.
What are the two deployment options in the DEPA draft?
Under Option 1, data providers anonymise their own data and developers can download it or train in a secure enclave. Under Option 2, the exchange anonymises raw data and access happens only inside the enclave.
Does DEPA for AI Training require individual consent?
The draft says individual consent of the kind used by India’s Account Aggregator system is impractical for pooled training data. It relies on de-identification, contracts, enclaves and differential privacy instead.
Who certifies AI models trained under DEPA?
The draft says a competent government authority should verify trained models as fit for use and can suspend or revoke non-compliant activity.
