What does data poisoning do?

• A. Injects harmful or biased data into training datasets to corrupt model outputs
• B. Reveals model weights to attackers
• C. Improves generalization of models
• D. Increases dataset diversity without harm

Answer: (A)

Data poisoning happens when an attacker contaminates the training data so the model learns incorrect patterns. By injecting harmful or biased examples into the training set, the model’s decision boundaries shift, causing corrupted outputs, misclassifications, or even hidden triggers (backdoors) that activate under specific inputs. This undermines the model’s reliability and trust, especially in systems that rely on continuously learning from user-supplied or open data. It’s not about exposing model weights, nor about genuinely improving generalization or simply increasing diversity; those outcomes would be constructive, while data poisoning intentionally steers the model toward faulty or biased behavior.