Which factor most directly contributes to training data leakage?

• A. Weak access controls.
• B. Model architecture complexity.
• C. Overfitting during training.
• D. Excessive data augmentation.

Answer: (A)

Training data leakage happens when training data is exposed to someone who should not have access. The most direct contributor is weak access controls, because if authentication, authorization, and data protection are insufficient, the data can be read, copied, or exfiltrated by unauthorized users or compromised systems, causing leakage regardless of how the model is trained. Strengthening access controls, enforcing least privilege, and auditing data access are key to preventing this risk.

While other factors relate to how a model learns or handles data, they do not create an exposure path as directly. Model architecture complexity affects learning capacity but not the presence of an accessible data channel. Overfitting can increase the risk of the model memorizing and potentially revealing training data through outputs, but the leakage route hinges more on who can access the data in the first place. Excessive data augmentation expands the dataset without inherently creating a leakage channel.