MeitY Floats New Privacy-Led Framework To Unlock Data For Training AI Models

DEPA for AI training is a proposed consent-based framework from the IT ministry, or MeitY. It would help startups, businesses and academic institutions access large-scale, India-specific datasets for AI development. Its central challenge is balancing useful data access with citizens’ privacy and compliance with the DPDP Act.

The framework combines technical and legal safeguards. Organisations could use confidential clean rooms to train models without exposing underlying datasets. Differential privacy would reduce the risk of identifying individuals. Electronic contracts would define approved uses, security duties and data protections, while zero-trust security would authenticate and authorise every access request.

The framework is still proposed, so its full effect depends on implementation. If adopted, it could make data sharing more practical for researchers and businesses. It also introduces checks for bias, fairness, security, safety and explainability before models move to pilot deployment or wider scaling.

Training Data Providers, or TDPs, are organisations that make datasets available. The article gives banks, hospitals and government agencies as examples. They contribute the underlying information, after anonymising and encrypting it, and register it on a data exchange platform.

Training Data Consumers, or TDCs, are the users seeking data for particular AI projects. They include startups, enterprises, researchers and individuals. They request datasets for defined use cases, but do not automatically receive unrestricted access to the raw information.

Data Collectives, also called Data Collabs, act as organising intermediaries. They aggregate, standardise and organise information from several sources. They also define data-sharing policies and governance rules. Together, these roles create a controlled chain from data contribution to model development, with contracts and security checks governing access.

The proposed architecture uses confidential clean rooms, which are secure computing environments for joint data processing and AI model training. These environments allow participating organisations to collaborate without exposing their underlying datasets to one another. This matters when several parties want the benefits of combined data but cannot or should not transfer raw records.

For example, a bank and a hospital could contribute protected datasets to a controlled training environment. A Training Data Consumer could request access for a stated use case, and an electronic contract system would check the request against sharing policies. Training would happen inside the secure environment, rather than by handing the consumer raw files.

Differential privacy adds another layer by helping ensure that individual-level information cannot be identified or extracted from aggregate data. Encryption, access controls, continuous monitoring and auditing support the clean room. The resulting model would still face independent checks before pilot deployment or scaling.

Consent establishes the basis for sharing data under the proposed framework. A Training Data Consumer requests a dataset for a specific use case, rather than receiving open-ended access. The request is checked against the provider’s data-sharing policies and governance rules before approval.

Electronic contracts make those conditions machine-readable and enforceable. They can specify permitted uses, data protection duties and security requirements. Providers would first anonymise and encrypt datasets, then register them with metadata and cryptographic identifiers for traceability. This creates a record of what data is involved and how it may be used.

Zero-trust security assumes every access request must be authenticated and authorised. Encryption, continuous monitoring and auditing help protect the data throughout the process. The framework also proposes ongoing monitoring, periodic reviews and output traceability, linking model results back to the data and contracts used in training. These controls support accountability after deployment.

AI developers need broad datasets because models learn from the patterns and examples present in their training material. India-specific data can better represent the country’s users, institutions, languages, services and conditions than datasets collected elsewhere. Smaller datasets may leave important groups or situations underrepresented, while foreign datasets may not match local realities.

For example, a model developed for Indian healthcare or public services could benefit from relevant data supplied by hospitals or government agencies. Under the proposed framework, providers would anonymise and encrypt datasets, while consumers would request access for defined purposes. Secure training would keep raw data inaccessible to consumers.

The article’s stated aim is to give startups, businesses and academic institutions access to large-scale, India-specific datasets. That access is not unrestricted: contracts, differential privacy and zero-trust controls would apply. Independent checks for bias, fairness, security, safety and explainability are also proposed before deployment.

If the framework works as intended, Indian startups, researchers and businesses could gain access to larger, more relevant datasets for developing AI models. The article says this could foster collaboration and help organisations train models without directly transferring the underlying data. That may reduce a major barrier for groups that lack large private datasets.

A startup could request data from a bank, hospital or government agency for a defined project. A Data Collective could organise information and set governance rules, while an electronic contract checked the request. Training would occur in a secure environment using privacy-preserving methods. The resulting model would need independent checks before pilot use or scaling.

Citizens could benefit from stronger privacy protections and traceability of model outputs back to training data and contracts. However, the framework remains proposed. Its impact will depend on implementation, continuous monitoring, periodic reviews and the effectiveness of its safeguards after deployment.

In general, AI models learn by examining many examples and adjusting internal parameters to recognise recurring patterns. The examples may contain text, images, records or other information. A model’s behaviour therefore reflects the data used to train it, including which people, situations and viewpoints appear frequently or are missing.

If training data is biased, a model may repeat or amplify unfair patterns. If it is incomplete, the model may perform poorly for groups or situations that were underrepresented. Poor protection creates a different risk: personal information may be identified, extracted or exposed. These risks explain why the proposed framework includes differential privacy and secure training environments.

The article also proposes independent checks for bias, fairness, security, safety and explainability. After deployment, Training Data Consumers would remain subject to monitoring, reviews and traceability. These measures cannot replace careful data governance, but they aim to support safer and more accountable AI development.