India’s MeitY unveils guide to train AI without sharing data

India’s Ministry of Electronics and Information Technology unveiled a handbook in Bengaluru on using the Data Empowerment and Protection Architecture, or DEPA, for AI training. The central problem is simple: institutions may want to cooperate on AI, but they do not want to surrender control of their datasets.

The handbook combines contracts, controlled computing environments, privacy safeguards, and data provenance. Together, these tools can record where a dataset came from, what permissions cover it, and how it is used during training. That creates a structured way to collaborate without treating data as something that must freely circulate.

The handbook was developed with support from the iSPIRT Foundation. At its launch, officials described DEPA as a way for models to move to data instead of data moving between institutions. The article does not report a specific deployment or measured result, but the framework is intended to support safer, permission-based AI collaboration.

DEPA stands for Data Empowerment and Protection Architecture. NITI Aayog has described it as a secure, consent-based architecture for data sharing, and MeitY hosts a DEPA document on its website. In this handbook, DEPA is applied to collaborative AI training.

The approach does not simply open a database to every participant. Contracts can define allowed uses, while controlled computing environments limit how work happens around the data. Privacy safeguards protect sensitive information, and data provenance records the dataset’s source, permissions, and use during training. These pieces help institutions cooperate without giving up overall control.

The article presents DEPA as a framework, not as a guarantee that every project will be safe automatically. Participants still need suitable agreements, technical controls, and permissions. MeitY’s handbook explains how those elements can work together, while officials highlighted the possibility of bringing models to data rather than transferring data elsewhere.

Instead of copying a dataset to a central location, an AI model can be sent to the institution that holds the data. The model then works inside a controlled computing environment, where access and permitted uses can be governed. This is the core idea highlighted by principal scientific adviser Ajay Kumar Sood.

For example, several institutions could allow an approved training process to run within their own environments. Contracts would set the rules, privacy safeguards would protect the data, and provenance records would track its source, permissions, and use. The data stays under the institution’s control rather than being transferred between participants.

The article does not describe the exact software, model architecture, or outputs used in such a process. It does establish the governing approach: move models to data, not data between institutions. In practice, the arrangement would depend on each participant’s permissions, computing setup, and compliance with the agreed safeguards.

The article does not name a particular data owner, cloud operator, or permission administrator for every project. It says institutions can work together while keeping control of their data. That means the participating data-holding institutions are the clearest identified controllers in the framework’s description.

The rules are supported by several mechanisms. Contracts can establish what participants may do. Controlled computing environments can restrict where and how training runs. Privacy safeguards protect the information, while data provenance tracks the dataset’s source, permissions, and use. These controls work together rather than placing all authority in one person.

MeitY launched the handbook, and the iSPIRT Foundation supported its development. Those roles do not mean either organization automatically owns participants’ data. The article also names S Krishnan and Ajay Kumar Sood as officials involved in the launch and discussion, but it does not say they personally control project data or permissions.

The article does not state how much data AI training can require. In general, training can involve anything from modest datasets to extremely large collections, depending on the model and task. The important point here is not a specific volume, but the difficulty of moving data between institutions.

Large transfers can take substantial time, computing capacity, and network resources. Sensitive datasets add risks because copying them creates more places where unauthorized access, accidental disclosure, or misuse could occur. Transfers can also make it harder to confirm whether every use matches the original permission.

DEPA’s proposed response is to reduce unnecessary movement. Contracts define permitted use, controlled environments constrain access, privacy safeguards protect information, and provenance records its source and permissions. The article does not quantify these risks or report a particular incident. It presents moving the model to the data as an alternative to repeatedly transferring datasets.

If institutions can jointly train models without directly sharing raw data, they may be able to combine the value of separate datasets while retaining local control. This could matter for organizations handling sensitive or restricted information. The article’s framework is designed to make that cooperation possible through permissions, contracts, and technical safeguards.

A hospital, government agency, or company could keep its dataset in a controlled computing environment. An approved model would work with the data there, while provenance tools record its source, permissions, and use. Privacy protections would help limit exposure, and contracts would define the collaboration’s boundaries.

The article does not report a hospital, agency, or company using DEPA this way, nor does it promise a particular improvement in model quality. Its stated implication is broader: institutions may collaborate on AI training without moving data between them. Whether that works depends on implementation, permissions, safeguards, and the participating institutions’ agreements.

Training an AI model is the process of using examples to adjust a model so it can perform a task. The examples provide information about patterns, relationships, or signals that the model needs to learn. Without relevant data, there is little basis for making those adjustments. This general explanation goes beyond the article’s brief reference to AI training.

In the article’s setting, institutions may contribute access to data without sending the raw datasets elsewhere. A model can move into a controlled computing environment, where approved training uses the institution’s data. Contracts, privacy safeguards, and provenance records help govern that process and document what data was used.

The article focuses on data control, not on algorithms, labels, model size, or training steps. It therefore does not specify how a model learns or how its performance would be measured. It does show why the training process creates a governance challenge: useful training requires data, while institutions may need to restrict how that data is accessed and shared.