Training Data Leakage: When Models Remember Too Much | Quiz

Why: When leakage is discovered the incident response plan should be activated including regulatory notification assessment – under GDPR the 72-hour clock starts immediately upon discovery. Context: You need to assess whether the leaked data triggers mandatory breach notification requirements and begin containment. Simply adding filters or retraining takes time you may not have. Remember: Discovery triggers the clock – activate incident response immediately.

Why: Fine-tuning on patient data creates high memorization risk – the model could leak patient information in responses to unrelated queries causing HIPAA violations and mandatory breach notifications. Context: A real 2024 incident involved exactly this scenario resulting in HIPAA fines and severe reputational damage. Healthcare data requires explicit safeguards beyond standard model training practices. Remember: Healthcare fine-tuning without safeguards leads to HIPAA violations.

Why: Canary tokens are fake but realistic-looking sensitive data inserted into training data – if they appear in model outputs you have proof of memorization and leakage. Context: This technique provides early warning before real sensitive data surfaces. Examples include synthetic API keys or fictional employee records that look authentic. Remember: Canary tokens are tripwires that prove memorization happened.

Why: Fine-tuning on a small unique private dataset causes those distinctive data points to become over-represented and extremely prone to verbatim leakage. Context: Fine-tuned models often memorize their fine-tuning data more strongly than the original base model memorized its training data. This is the opposite of what many organizations assume. Remember: Fine-tuning increases memorization risk – it does not decrease it.

Why: Training data becomes part of the model permanently – even if you delete the original training files the information persists in the model parameters with no undo button available. Context: This is why prevention is essential – once sensitive data enters a trained model removing it is an unsolved research problem. Your training data does not just influence the model – it becomes the model. Remember: No undo button exists for learned information.

Why: Leakage involves models reproducing real sensitive information that actually existed in training data while hallucination involves generating false or nonsensical information. Context: This distinction matters for risk assessments – leakage is a privacy and security failure while hallucination is a fidelity and utility failure. They require different mitigation strategies. Remember: Leakage is real data escaping – hallucination is fake data appearing.

Why: Training data leakage occurs when AI models unintentionally memorize and reproduce sensitive information from their training data in their outputs. Context: This includes PII and confidential documents and copyrighted content – creating privacy violations and compliance risks. It is an inherent property of how neural networks learn – not a bug to fix. Remember: Models remember training data and can reveal it in outputs.