Re identification risks from simulated datasets

Introduction

Re-identification risks from simulated datasets have become an important legal and ethical concern in data protection, artificial intelligence, biotechnology, healthcare research and scientific data management. Simulated datasets are artificially generated datasets designed to reproduce the statistical characteristics, patterns and relationships of real-world data without directly reproducing the original records. They are frequently used for medical research, machine learning, clinical trials, algorithm development and scientific experimentation.

Although simulated datasets are generally considered safer than original personal datasets, they may still create privacy risks. If generated data retain distinctive characteristics, rare combinations or statistical patterns associated with identifiable individuals, external users may potentially connect the simulated records with real persons. This process is known as re-identification.

Re-identification risks are particularly significant in healthcare and biotechnology because simulated datasets may contain information resembling genetic profiles, medical histories, disease patterns and demographic characteristics. Therefore, organisations must ensure that synthetic data generation does not unintentionally expose confidential information.

Meaning and nature of re-identification risks

Re-identification refers to the process of identifying an individual from information that has been anonymised, pseudonymised or otherwise modified. In simulated datasets, re-identification may occur when artificially generated records retain sufficient information about actual individuals.

For example, a synthetic healthcare dataset may reproduce unusual combinations of age groups, medical conditions, geographical characteristics and treatment histories. When these combinations are compared with publicly available information or external databases, an individual may become identifiable.

The major re-identification risks include:

Membership inference: Determining whether a particular individual's original information was included in the training dataset used to generate synthetic data.

Attribute disclosure: Discovering sensitive information about an individual through statistical patterns preserved in simulated records.

Linkage attacks: Connecting simulated records with external databases containing identifiable information.

Memorisation risks: Occurrence of situations where machine learning models reproduce specific information from their training datasets.

Inference attacks: Deriving confidential personal characteristics from relationships and patterns contained in generated datasets.

These risks demonstrate that synthetic data cannot automatically be treated as completely anonymous.

Legal framework governing simulated datasets in India

India's Digital Personal Data Protection Act, 2023 provides an important framework for protecting digital personal data. Where simulated datasets contain or enable access to identifiable personal information, relevant data protection obligations may arise depending on the nature of the information and processing activities.

The Act emphasises lawful processing, appropriate safeguards and accountability of Data Fiduciaries. Organisations handling original personal information for synthetic dataset generation must consider the purpose and legal basis of processing, security safeguards and applicable obligations.

The Information Technology Act, 2000 also provides relevant protections. Section 43 addresses unauthorised access and specified forms of damage involving computer systems, while Section 66 deals with certain acts described under Section 43 when committed dishonestly or fraudulently. Section 72 addresses breach of confidentiality and privacy in specified circumstances.

In healthcare and biotechnology, the ethical guidelines issued by the Indian Council of Medical Research (ICMR) are relevant to research involving human participants and associated personal information. Researchers should ensure confidentiality, responsible data management and appropriate ethical oversight.

Internationally, the General Data Protection Regulation (GDPR) is significant. Recital 26 explains that information is anonymous only where individuals are not identifiable, considering reasonably likely means of identification. Article 4 distinguishes personal data from anonymised information and establishes requirements relevant to pseudonymisation.

Technical and institutional measures to reduce re-identification

Organisations should adopt a combination of technical, organisational and legal safeguards.

1. Differential privacy

Differential privacy introduces carefully calibrated statistical noise into datasets or query results. It reduces the likelihood that information about a particular individual can be inferred from the output.

2. Data minimisation

Only information necessary for the intended research or analytical purpose should be used during synthetic data generation.

3. Membership inference testing

Organisations should evaluate whether attackers can determine whether particular individuals were included in the original training dataset.

4. Regular privacy assessments

Privacy impact assessments should examine the possibility of linkage attacks, attribute disclosure and unintended memorisation.

5. Access control and encryption

Sensitive original datasets, model parameters and generated outputs should be protected through appropriate authentication, encryption and access restrictions.

6. Independent auditing

Independent experts should periodically evaluate the privacy characteristics of synthetic datasets before their distribution or commercial use.

Important case laws

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

The Supreme Court of India recognised the right to privacy as a fundamental right under Article 21 and other provisions of the Constitution. The judgment established that informational privacy is an important element of individual autonomy and dignity.

Its principles are relevant to simulated datasets because organisations must consider whether data generation and subsequent processing create risks to individuals' privacy, even when original identifiers are removed.

2. District Registrar and Collector v. Canara Bank (2005)

The Supreme Court examined privacy protections concerning access to confidential financial records. The decision reinforces the importance of protecting sensitive information from unjustified disclosure.

Its principles are relevant when simulated financial or healthcare datasets preserve confidential characteristics of original records.

3. R. Rajagopal v. State of Tamil Nadu (1994)

The Supreme Court recognised the right to privacy and discussed limitations on publishing private information without consent, subject to recognised legal exceptions.

The judgment is relevant to synthetic datasets because the removal of direct identifiers does not necessarily eliminate privacy concerns where an individual can still be recognised through associated information.

4. Google Spain SL v. Agencia Española de Protección de Datos (2014)

The Court of Justice of the European Union examined personal data protection in relation to search engine processing and the rights of individuals concerning information about them.

Although the case did not directly concern simulated datasets, it illustrates the broader legal importance of controlling the processing and accessibility of information that can be associated with identifiable individuals.

5. 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 enabling identification was available through another party.

The judgment is particularly relevant to re-identification because it demonstrates that information need not independently identify an individual to qualify as personal data when identification is reasonably possible using additional information.

Challenges and regulatory concerns

The principal challenge is determining whether simulated data are genuinely anonymous or merely difficult to identify. Organisations may incorrectly assume that generated datasets are risk-free because they do not contain names, addresses or identification numbers.

Another challenge involves technological advancement. Machine learning systems can analyse large volumes of information and identify relationships that were previously difficult to detect. Consequently, datasets considered sufficiently protected at the time of generation may become vulnerable as external databases and analytical capabilities develop.

Cross-border research collaborations also create additional concerns because different jurisdictions apply different standards of privacy protection, consent and data governance.

Regulators should therefore encourage periodic reassessment of synthetic datasets, transparent documentation of generation methods and independent privacy testing.

Conclusion

Re-identification risks from simulated datasets demonstrate that artificial data generation does not automatically guarantee complete privacy protection. Synthetic datasets may preserve identifiable characteristics, statistical relationships or sensitive information from original records. Indian constitutional privacy jurisprudence, particularly the Puttaswamy judgment, provides an important foundation for protecting individuals against inappropriate information processing.

Differential privacy, data minimisation, independent audits, access controls and regular privacy assessments can reduce these risks. Ultimately, responsible synthetic data governance requires organisations to balance scientific innovation and research efficiency with confidentiality, individual privacy and legal accountability.