What term describes an attack that injects malicious samples into a machine learning training dataset to corrupt the model's behavior?

• A. Data Poisoning
• B. Adversarial Training
• C. Data Tampering
• D. Model Inversion

Answer: (A)

This question tests understanding of training-time attacks on machine learning, specifically how corrupting the training data can alter a model’s behavior. When an attacker injects malicious samples into the training dataset to steer the learning process, degrade accuracy, or create a hidden backdoor, that is called data poisoning. It focuses on contaminating the data the model learns from, rather than manipulating inputs after the model is trained.

Adversarial training is a defensive or strengthening technique that uses adversarial examples during training to make models more robust, not an attack. Data tampering is a broader term for altering data in general, but data poisoning specifically refers to injecting harmful samples into the training data to influence the model. Model inversion involves trying to reconstruct training data or private information from the model’s outputs, not poisoning the training data.