Which measure would most effectively reduce the risk of training data leakage?

• A. Strengthen access controls and data-handling policies.
• B. Increase model size to improve privacy by obscurity.
• C. Rely on automated data pruning without governance.
• D. Distribute data across multiple clouds without encryption.

Answer: (A)

Reducing training data leakage comes down to controlling who can access data and how it’s used, backed by clear data governance. Strengthening access controls and data-handling policies directly limits exposure: it enforces least privilege, strong authentication and authorization, regular monitoring, and defined rules for data classification, masking or anonymization, retention, and transfers. When these policies are in place, only authorized personnel can access sensitive data, and there are auditable safeguards that prevent misuse or accidental exposure, which is the core way to minimize leakage from model training.

The other options don’t provide this solid governance foundation. Simply increasing model size does not protect data; it can even increase the risk by enabling the model to memorize more training data. Relying on automated data pruning without governance leaves missing controls and oversight, so sensitive information could still remain in the training data. Distributing data across multiple clouds without encryption adds exposure and neglects essential protective measures; without encryption and proper key management, broader data distribution heightens leakage risk rather than reducing it.