Data Extraction
Retrieving Sensitive Data from Training Datasets
Description
Data extraction attacks target a model’s training data, enabling an adversary to recover sensitive or proprietary information without directly accessing its storage. These attacks exploit a model’s unintended memorization (overfitting) to reveal whether specific records were used (membership inference) or to reconstruct entire training samples. Data extraction attacks often occur post-training when models are exposed via public APIs or deployed on edge devices. Extracted data can be used to clone the system, reconstruct proprietary products, or resell exclusive datasets (e.g., paid financial data).
Difficulty Level — High
Requires advanced technical skills and understanding of the model type and data complexity. Attack difficulty depends on model type; for example, diffusion models and overparameterized neural networks tend to memorize more, increasing extractability.
Attack Consequence — High
Can trigger regulatory and legal action, cause loss of proprietary data and competitiveness, damage reputation, and enable further attacks using the recovered data.