What is label poisoning and how can it affect model quality?

• A. Incorrect/targeted labels injected during labeling; leads to degraded performance or biased outcomes.
• B. A method to speed labeling.
• C. A method to encrypt labels.
• D. A process to anonymize training labels.

Answer: (A)

Label poisoning is when incorrect or malicious labels are introduced into the training data during labeling. This corrupts the supervisory signal the model relies on to learn mappings from inputs to outputs. Because the model learns from mislabeled examples, the decision boundary shifts away from the true patterns, leading to degraded performance on new, clean data. The overall accuracy drops and the model may be poorly calibrated, making predictions less reliable. If the poisoning is targeted, it can cause systematic errors for specific inputs or scenarios, effectively creating a backdoor that triggers misclassification. To protect model quality, use careful labeling workflows, enable multiple annotators and audits, validate labels, and apply training methods that are robust to some label noise.