AI/ML model poisoning involves adversaries manipulating training data to degrade model performance. Which of the following is a mitigation strategy?

• A. Mitigations include data validation, robust training, monitoring, and model hardening.
• B. Only using unsupervised learning prevents poisoning.
• C. Stop using data entirely.
• D. Hiring more data scientists increases poisoning risk.

Answer: (A)

Defending against training-time data poisoning relies on protecting the data pipeline and making learning robust to imperfect data. Data validation catches suspicious or mislabeled samples before they influence the model, helping to stop poisoned data from entering the training set. Robust training techniques reduce sensitivity to outliers or maliciously crafted points, so even if some data slips through, the model isn’t easily hijacked by a small amount of corrupted information. Monitoring the training process and model behavior provides early warning of unusual patterns or shifts in data distribution, allowing an intervention before harm accumulates. Model hardening—through defenses like adversarial training, differential privacy, and secure data handling—further reduces how much a manipulated example can affect the model’s parameters and outputs.

This combination is effective because it addresses multiple stages of the threat: preventing bad data from influencing training, making the learning process itself more resistant to corrupted data, and keeping an eye on signals that something is being tampered with. In contrast, relying on unsupervised learning doesn’t inherently stop poisoning, stopping data usage entirely is impractical, and simply hiring more data scientists without stronger data safeguards can even raise risk by widening exposure to tainted data.