MeitY’s DEPA Framework for AI Training for Responsible Data Collaboration
The strength of Artificial Intelligence (AI) is heavily dependent on the data on which it is trained. The diversity of this data directly determines an AI model’s capabilities, fairness, and reliability. In this context, India stands at a critical juncture, generating unprecedented volumes of digital data across sectors like healthcare, finance, and agriculture. But this large volume of data sits divided across domains, safeguarded behind different regulatory barriers, including-
- legal and regulatory uncertainty around data sharing,
- technical incompatibilities between systems, and
- lack of institutional understanding on best practices for data security and privacy.
It eventually leads to data deadlock, especially when the traditional techniques of anonymisation and model training are often found to be insufficient in maintaining guarantees of privacy and confidentiality. Therefore, even where data sharing could support innovation, the associated risks may discourage institutions from participating.
To address this challenge, the Ministry of Electronics and Information Technology (MeitY) has released the Data Empowerment and Protection Architecture (DEPA) Framework for AI Training. The framework seeks to establish a techno-legal mechanism through which organisations can collaborate in training AI models without directly exchanging or exposing their underlying datasets.
Why AI Training Requires a Different Approach to Data Sharing
The AI models use two distinct processes for functioning- AI inference and AI training.
AI inference uses an already trained AI model to generate a prediction or decision based on a particular input. For instance, a bank may use an existing credit-scoring model to assess an individual’s financial information and determine eligibility for a loan.
AI training, on the other hand, develops the intelligence of the model itself. It analyses large datasets to identify patterns, correlations and trends. The value lies not in any single individual’s information but in the collective insights that emerge from analysing data across a larger population. For example, let us consider a situation where three retail businesses want to jointly develop an AI model to predict consumer purchasing behaviour. Each business holds valuable transaction information. Combining insights from these datasets could help develop a more accurate model of market trends. However, sharing raw customer data could expose personal information, sales figures, customer relationships and commercially sensitive business information.
Traditional anonymisation may not completely address these concerns. With sufficiently large and interconnected datasets, individuals may sometimes be re-identified by combining information from different sources. Excessive anonymisation may also reduce the usefulness of the data for AI training. So, even when the organisations need to collaborate to improve AI models, the risks associated with such collaboration itself can create dilemma for them.
DEPA as a Techno-Legal Solution for AI Training
The DEPA Framework for AI Training seeks to address this dilemma by enabling data analysis without requiring the underlying datasets to be transferred between participating organisations.
It builds on the broader framework which has already been operationalised in the financial sector through the Account Aggregator framework. However, the two approaches serve different purposes.
Under the existing DEPA inference model, an individual can provide explicit, granular and revocable consent for their financial information to be shared between institutions for a specific purpose. For example, an individual may consent to sharing bank statements with a lending institution to support a loan application. While DEPA for AI Training enables organisations to collaborate on large-scale data analysis and AI model development while protecting individual privacy and institutional interests.
The framework is based on ADEPTS, i.e., the six guiding principles, namely:
- Accountability: Ensuring that participants remain accountable for their roles and responsibilities.
- Democratization: Enabling wider participation in responsible data collaboration and AI innovation.
- Evolvability: Allowing the framework to adapt to changing technologies and sector-specific requirements.
- Privacy: Protecting individuals’ information during data processing and AI training.
- Transparency: Making data-sharing arrangements and permitted uses clear to participants.
- Security: Establishing safeguards against unauthorised access, disclosure and misuse.
Together, these principles seek to create an environment where data collaboration is governed by both technological safeguards and enforceable institutional arrangements.
Further, the framework envisages collaboration between two principal participants- Training Data Providers (TDPs) and Training Data Consumers (TDCs)-
- TDPs are institutions such as hospitals, banks, public authorities and businesses that possess valuable datasets.
- TDCs include AI developers, start-ups, researchers and enterprises that require access to data insights for developing AI models.
Their interaction is supported by four key principles:
- Data Exchange Platform: Facilitates the coordination and exchange of information required for data collaboration.
- Electronic Contracts: Establish the terms and conditions governing data use, including the permitted purposes, obligations and restrictions applicable to participating organisations.
- Secure Data Environment (SDE): Provides a controlled environment in which computations and analysis can take place without exposing the underlying raw datasets to the data consumers.
- Differential Privacy: Introduces privacy-protective techniques during data analysis and model training, reducing the risk that information relating to a particular individual can be identified or extracted from the resulting outputs.
Another important feature of this framework is that the responsibility for anonymisation rests with the TDPs. The data remains protected within the Secure Data Environment, while the AI training process uses privacy-preserving techniques to derive aggregate insights. This approach is intended to reduce the risks associated with conventional data transfers and make collaboration more secure and auditable.
New Approach for Responsible Data Collaboration
The DEPA Framework could create new opportunities for businesses to participate in collaborative AI development. However, it also brings data governance and regulatory compliance considerations to the forefront. So, the organisations must ensure the following:
1. The use of data for AI training does not automatically remove an organisation’s data protection obligations. Organisations need to careful about the obligations for data collection, processing and sharing arrangements under the Digital Personal Data Protection Act, 2023.
2. Data-sharing agreements need stronger contractual safeguards. Electronic Contracts under the framework provide an important mechanism for defining the rights and responsibilities of participating organisations which contribute data to develop a common AI model.
3. Organisations must evaluate whether the proposed collaboration adequately protects trade secrets, intellectual property and other confidential business assets.
4. Privacy-enhancing technologies can reduce risks, but they do not replace organisational accountability. Legal, information security, data protection and business teams will need to work together rather than approach AI training as an exclusively technical exercise.
5. Organisations need to evaluate sector-specific requirements alongside general data protection and cybersecurity obligations before participating in collaborative data initiatives. Because, a bank, hospital, insurance company and retail business may all participate in AI training, but the regulatory restrictions applicable to their underlying datasets may differ significantly.
This DEPA Framework for AI Training represents an important evolution in India’s approach to data governance. Its success, however, will depend on effective implementation. The reliability of Secure Data Environments, the practical effectiveness of Differential Privacy, the enforceability of electronic contracts and the accountability of participating institutions will determine whether the framework can deliver its intended benefits.
The implementation costs, institutional readiness, governance standards and the allocation of risks where data misuse or security incidents occur are also crucial factors. The role of Self-Regulatory Organisations in developing sector-specific standards and maintaining accountability will be important in this regard.
In essence, responsible AI innovation will require more than access to data or sophisticated technology. It will require trust, transparency, accountability and a clear understanding of the legal boundaries governing data use, while keeping regulatory compliance and individual privacy at the heart of the organisations’ AI strategies.
