Data anonymization for training models

Introduction

Data anonymization refers to the process of removing or transforming personally identifiable information from datasets so that individuals cannot be identified directly or indirectly. It is increasingly important in artificial intelligence (AI), machine learning and biotechnology because these technologies require large volumes of data for training and testing models.

Training datasets may contain names, addresses, medical records, genetic information, financial details, photographs and behavioural patterns. If such information is improperly disclosed, individuals may face privacy violations, discrimination or identity theft.

Anonymization attempts to reduce these risks while allowing organisations to use datasets for research, innovation and model development. However, anonymization must be effective because removing names alone does not necessarily prevent identification.

Legal framework in India

1. Digital personal data protection act, 2023

The Digital Personal Data Protection Act, 2023 (DPDP Act), establishes India's framework for processing digital personal data. Its provisions are relevant when organisations collect personal information for AI model training.

Section 3 defines the scope of the Act, while Section 4 establishes requirements concerning lawful processing of digital personal data. Section 6 deals with consent, and Section 8 specifies obligations of data fiduciaries.

Information that has been irreversibly anonymized so that no individual can be identified may fall outside the definition of personal data. However, pseudonymized information remains personal data where an individual can still be identified using additional information.

The legal treatment also depends on the commencement of relevant provisions under the Act.

2. Information technology act, 2000

The Information Technology Act, 2000, provides additional protection for digital information.

Section 43A historically addressed compensation for failure by a body corporate to maintain reasonable security practices while handling sensitive personal data or information. Its present operation must be considered alongside the commencement and enforcement of the DPDP framework.

Section 72 addresses breach of confidentiality and privacy by persons who obtain access to information through powers conferred under the Act.

These provisions demonstrate the importance of protecting personal information during digital processing and storage.

3. Information technology rules, 2011

The Information Technology (Reasonable Security Practices and Procedures and Sensitive Personal Data or Information) Rules, 2011, established requirements concerning sensitive personal information, including financial, medical and biometric information.

Their application must be assessed in light of subsequent legislative changes and the DPDP framework.

Methods of data anonymization

Organisations use several technical methods to reduce identification risks in AI training datasets.

1. Data masking: Direct identifiers such as names, telephone numbers and email addresses are replaced with artificial values.

2. Generalisation: Specific information is converted into broader categories. For example, an exact age may be replaced with an age range.

3. Data aggregation: Individual records are combined into statistical groups so that information about particular individuals becomes difficult to identify.

4. Differential privacy: Mathematical techniques introduce controlled statistical noise into datasets or model outputs, reducing the possibility of identifying individual contributors.

5. Synthetic data generation: Artificial datasets are generated to reproduce useful statistical patterns without directly reproducing original personal records. However, synthetic data is not automatically anonymous if it reproduces identifiable information.

Important case laws

1. Justice K.S. Puttaswamy v. Union of India (2017)

The Supreme Court recognised privacy as a fundamental right under Article 21 of the Constitution.

The judgment established that informational privacy is an essential part of personal liberty and that interference with privacy must satisfy constitutional requirements.

Relevance: Organisations training AI models using personal datasets must respect privacy and ensure that data processing has a lawful basis. Anonymization can reduce privacy risks, although it does not automatically establish that every processing activity is lawful.

2. K.S. Puttaswamy v. Union of India (Aadhaar case, 2018)

The Supreme Court examined the constitutional validity of the Aadhaar framework and considered issues involving biometric information, identification and informational privacy.

The judgment recognised the importance of safeguards against misuse of personal information.

Relevance: AI training datasets containing biometric and identity-related information require strong safeguards. Removing direct identifiers may be insufficient where unique biometric characteristics can still identify individuals.

3. Google Spain SL v. AEPD and Mario Costeja González (2014)

The Court of Justice of the European Union examined the rights of individuals concerning personal information appearing in search engine results.

The judgment recognised the importance of individual control over personal information in the digital environment.

Relevance: Although the case did not directly concern AI anonymization, it demonstrates the importance of protecting identifiable personal information when digital systems process and organise large datasets.

4. Breyer v. Bundesrepublik Deutschland (2016)

The Court of Justice of the European Union considered whether dynamic IP addresses could constitute personal data where additional information could potentially identify an individual.

Relevance: The decision demonstrates that information should not be treated as anonymous merely because direct identifiers have been removed. The possibility of identification through additional information must also be considered.

Challenges in AI model training

Anonymization presents several practical and legal challenges:

Re-identification: Individuals may be identified by combining anonymized datasets with publicly available information.

Model memorisation: AI models may reproduce sensitive information contained in their training datasets.

Data utility: Excessive anonymization may reduce the accuracy and usefulness of training data.

Genetic information: Genetic characteristics may remain identifiable even after names and addresses are removed.

Third-party sharing: Organisations must ensure that external researchers and technology providers follow appropriate data protection safeguards.

Conclusion

Data anonymization is an important privacy protection measure in AI model training, biotechnology research and digital innovation. Indian law, particularly the DPDP Act, 2023, and constitutional principles recognised in the Puttaswamy judgments, provides an important foundation for responsible data processing. International decisions such as Breyer demonstrate that removing direct identifiers does not necessarily make information anonymous. Organisations should therefore combine technical methods, re-identification risk assessments, access controls and continuous monitoring to protect individuals while supporting responsible AI development.