General Data Protection Regulation (GDPR)

The process of permanently removing or altering personally identifiable information (PII) from a dataset so that individuals cannot be re-identified. For multimodal datasets, this may involve:

• Pixelation or blurring of faces and license plates in video and image data.
• Audio redaction to remove identifiable voices or background conversations.
• Text scrubbing to replace names, locations, and other identifiers with generic tokens. Crucially, GDPR considers anonymization a high bar; data must be irreversible. Simple pseudonymization, where a key exists to re-identify, is not sufficient for full GDPR exemption.

A rigorous mathematical framework for quantifying and limiting privacy loss when data is used in statistical analyses or machine learning. DP provides a provable guarantee that the inclusion or exclusion of any single individual's data does not significantly affect the output of an algorithm.

• Epsilon (ε) Parameter: The privacy budget; lower values mean stronger privacy guarantees but potentially noisier results.
• Noise Injection: Achieved by adding calibrated random noise (e.g., from a Laplace or Gaussian distribution) to query results or model gradients. DP is a powerful tool for GDPR-compliant dataset analysis and model training, as it allows for the extraction of aggregate insights while mathematically bounding the risk to any individual.

The documented history of a dataset's origin, ownership, transformations, and processing steps. For GDPR compliance, provenance is critical for demonstrating accountability and enabling the right to erasure.

• Lineage Tracking: Logging every operation performed on personal data, from ingestion through annotation, transformation, and model training.
• Consent Records: Linking data samples to the legal basis for processing (e.g., user consent forms, contract IDs).
• Impact Analysis: When an erasure request is received, a robust provenance system can identify all derived data and models that incorporated that individual's information, enabling complete deletion.

The overarching framework of policies, standards, roles, and processes that ensure the formal management of data. GDPR compliance is a primary driver of data governance programs. Key components include:

• Data Classification: Tagging data based on sensitivity (e.g., "Public," "Internal," "Confidential," "GDPR Personal Data").
• Access Controls & Role-Based Permissions: Enforcing the principle of least privilege for who can view or process personal data.
• Data Retention Policies: Defining and automating the deletion of data after its lawful purpose has expired.
• Data Protection Officer (DPO): The designated role responsible for overseeing GDPR compliance strategy.

A core GDPR mandate requiring that data protection measures are integrated into the development process of systems and business practices from the outset, not added as an afterthought.

• By Design: Architecting multimodal data pipelines to minimize data collection, implement anonymization early, and embed access logging.
• By Default: Ensuring system settings automatically provide the highest level of privacy (e.g., the strictest access controls are the default). For ML teams, this means evaluating the necessity of each data field collected, choosing algorithms that can work with encrypted or federated data, and building data minimization into annotation schemas.

While not explicitly named in GDPR text, the regulation's principles of fair and transparent processing create a strong link to fairness in AI. Models trained on personal data must not produce discriminatory effects.

• Bias Auditing: The systematic process of evaluating a dataset or model for unfair representations across demographic groups. Under GDPR, a biased model could be challenged as an unfair processing activity.
• Right to Explanation: Although limited, GDPR provides a right for individuals to obtain "meaningful information about the logic involved" in automated decisions. This pushes teams toward interpretable models and robust model cards that document performance across subgroups.