India’s Ministry of Electronics and Information Technology, working with the iSPIRT Foundation, has published a draft framework called DEPA for AI Training — a proposed techno-legal architecture for sharing aggregate, bulk datasets to train AI models across sectors without exposing raw records. The document positions the design as second-generation Digital Public Infrastructure, built for the age of AI.
It’s still a “Draft for Discussion,” so it represents the government’s plan rather than a live deployment. The draft also draws a hard boundary between this framework and DEPA Inference, the Account Aggregator (AA) consent system in finance that Sahamati orchestrates, which handles individual-level data for inferencing rather than model training.
Four pillars and two deployment options
The architecture rests on four technical pillars:
- A Data Exchange Platform – AI Kosh is named as the candidate – provides discovery, registration, and access.
- Electronic Contracts are machine-readable, cryptographically signed agreements that auto-enforce data-use terms.
- A Secure Data Environment (SDE) supplies hardware enclaves where training occurs without raw-data exposure, while differential privacy applies calibrated statistical noise governed by audited privacy budgets.
- Differential Privacy: The design follows the ADEPTS principles – which pair accountability and transparency with privacy and security – and it’s built to interoperate with existing national exchanges including AI Kosh and API Setu, plus the National Data and Analytics Platform.
Roles are split three ways. Training Data Providers – hospitals, banks, and public agencies – register, anonymise, and encrypt datasets. Training Data Consumers such as startups, researchers, and regulators train models, and commercial sector-specific Data Collabs aggregate data within their domains.
Two deployment options remain open in the draft: providers handle anonymisation themselves and consumers may train in the SDE or download data, or the exchange performs anonymisation and access happens only inside the enclave.
Training inside the secure enclave
The proposed workflow runs in five steps. Datasets are registered, encrypted, and cryptographically tagged or watermarked, then consumer requests are auto-validated against e-contract terms. Training is confined to the SDE, where raw data is never exposed, differential privacy is applied and logging is immutable.
Independent validation then covers bias and fairness audits alongside adversarial robustness testing and explainability checks – a discipline that parallels emerging practices in AI agent testing with realistic data – before deployment with continuous monitoring and traceability back to source data and contracts.
Security is specified as zero-trust, with encryption at rest, in transit, and in processing, plus real-time anomaly monitoring. The framework supports federated learning and secure multi-party computation, and pseudonymised keys stay with the originating agency.
Datasets are classified as open, registered, or restricted under the NDSAP Policy of 2012. An optional “AI Provenance Chain” distributed ledger would immutably record all data processing, and the draft claims alignment with the DPDP Act 2023, the EU AI Act, and HIPAA, among other regimes.
Federated learning and the economic case
The draft’s benchmark proof point is the Penn Medicine and Intel project, a 71-institution federation that trained a brain-tumor detection model through federated learning in a confidential clean room without sharing raw patient data. That model reached 99 percent accuracy relative to a hypothetical centrally trained model — evidence, the authors argue, that privacy-preserving training doesn’t sacrifice performance.
The macro framing ties the framework to AI’s projected $15 trillion addition to the global economy by 2030. India’s own stake is put at $500 billion of GDP within the same horizon. An RBI expert committee estimate puts unmet MSME credit at ₹20–25 trillion, and the draft argues that alternate-data AI underwriting could reach 400 million individuals and small businesses currently outside the formal credit system.
Shri Abhay Karandikar, Secretary of the Department of Science and Technology at the Government of India, said: “For too long, Indian researchers have faced a critical ‘data chasm’—possessing world-class talent, but often lacking access to the large-scale, high-quality, India-specific datasets required to solve our most pressing challenges.
“The DEPA for Training framework is a direct and powerful response to this chasm. We see it not as a regulatory document, but as a fundamental piece of our National Infrastructure.”
See also: OpenAI and Poland’s NASK partner on AI safety research
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including IoT Tech Expo and the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.