Data Protection and Cybersecurity in Artificial Intelligence

AI data protection covers the lawful and secure handling of information used by AI systems, while cybersecurity protects the systems and models themselves. Both require safeguards throughout the data lifecycle, from collection and training to inference, storage and deletion.

Foundations and Legal Frameworks

AI systems depend on large datasets and can process personal, financial, health or commercially sensitive information. A breach may expose data, enable misuse of a model or undermine the reliability of its outputs. Protection therefore involves technical controls, legal duties and clear accountability for decisions made across the system’s lifecycle.

• Digital Personal Data Protection Act, 2023: India’s law governs the processing of digital personal data. It establishes rights for data principals and duties for data fiduciaries, provides for consent and specified legitimate uses, and requires reasonable security safeguards. It also provides penalties for breaches and non-compliance. Its territorial scope includes certain processing outside India connected with offering goods or services to people in India.
• Information Technology Act, 2000: The Act gives legal recognition to electronic transactions and digital signatures and addresses cyber offences, including unauthorised access and data theft. Section 43A provides for compensation where a body corporate’s negligence in maintaining reasonable security practices causes wrongful loss or gain. The Act also supports protection of computer systems designated as critical information infrastructure.
• General Data Protection Regulation: The EU’s GDPR is an influential privacy framework. It provides rights including access, rectification, erasure and data portability, and includes data-breach notification and data-protection impact assessment requirements.
• NIST AI Risk Management Framework: This voluntary US framework supports trustworthy AI by addressing risks including privacy, fairness, safety and security. Its four functions are Govern, Map, Measure and Manage.

Privacy regulation establishes duties, but effective protection also depends on organisational capacity to identify risks, secure data and respond to incidents. Accountability mechanisms should clarify who is responsible for data handling, model deployment and remediation when a system causes harm or is compromised.

Cybersecurity Threats to AI Systems

AI systems face attacks on data, models and the infrastructure used to build or operate them. The effects can include compromised confidentiality, unreliable outputs, degraded performance and loss of proprietary technology.

• Data poisoning: Malicious or misleading examples are inserted into training data, distorting what a model learns and potentially producing biased or incorrect outputs.
• Adversarial evasion: Carefully modified inputs can cause a trained model to misclassify data during inference, such as an image altered so that it is incorrectly identified.
• Training-time poisoning and backdoors: An attacker can manipulate training data to introduce vulnerabilities or cause a model to behave in a particular way when it encounters a trigger.
• Model inversion and membership inference: Inversion attempts to reconstruct sensitive training information from a model’s outputs; membership inference seeks to establish whether a particular record was in the training dataset.
• Model extraction and supply-chain attacks: Repeated queries may help an attacker replicate a proprietary model. Vulnerable third-party libraries, frameworks or datasets can also introduce risks into AI development.

Technological Safeguards

Privacy-enhancing technologies can reduce exposure of raw data, while security controls must also protect models, infrastructure and information during processing. These methods have different purposes and may be combined according to the use case.

• Homomorphic encryption: Enables computation on encrypted data without first decrypting it, helping protect information processed by a cloud-based AI service.
• Federated learning: Trains a model across decentralised datasets, such as data held on user devices, without requiring raw data to be centralised. Model updates are aggregated instead.
• Secure multi-party computation: Allows several parties to jointly compute a result using private inputs while keeping those inputs confidential from one another.
• Differential privacy: Adds statistical noise to data or model outputs to limit what can be inferred about an individual record, while retaining analytical utility.
• Confidential computing: Uses hardware-based trusted execution environments (TEEs) to protect data in use, including from other software layers such as the operating system or hypervisor.
• Explainable AI: Makes model decisions more understandable to people, supporting scrutiny for bias or possible manipulation. Explainability complements, but does not replace, security controls.

Institutions and Accountability

• CERT-In: India’s nodal agency for responding to computer security incidents. It issues alerts and vulnerability notes and coordinates incident response.
• NCIIPC: The National Critical Information Infrastructure Protection Centre works to protect critical information infrastructure from cyber threats.
• Cyber Swachhta Kendra: CERT-In’s Botnet Cleaning and Malware Analysis Centre provides tools to detect and remove malicious software. Internationally, the G20, OECD and UN bodies contribute principles and guidance on responsible AI, including privacy and security.

Effective AI governance needs more than technical safeguards. Organisations must identify risks, assign responsibility, maintain security practices and respond to incidents. Privacy regulation, cybersecurity capacity and accountability mechanisms are therefore central to the trustworthy deployment of AI.

Key Prelims Takeaways

• DPDP Act, 2023: Governs digital personal data and requires data fiduciaries to implement reasonable security safeguards.
• IT Act, 2000: Addresses cyber offences; Section 43A concerns compensation for negligence in security practices.
• NIST AI RMF: Its functions are Govern, Map, Measure and Manage.
• AI threats: Include poisoning, evasion, model inversion, membership inference and model extraction.
• Privacy technologies: Federated learning avoids centralising raw data; homomorphic encryption permits computation on encrypted data.
• Indian institutions: CERT-In responds to cyber incidents, while NCIIPC focuses on critical information infrastructure.

Recent Context

On November 9, 2026, Google launched TPU 8tx and TPU 8i Max as additions to its eighth-generation TPU portfolio. TPU 8tx targets confidential inference and small-scale training, using hardware-level security to protect data and AI models during processing.