Which long-term strategy is described for eradicating data poisoning in AI systems?

• A. Selective retraining on clean data to force the model to forget poisoned data.
• B. Replacing the dataset entirely with random data.
• C. Archiving poisoned data and continuing with the same model.
• D. Increasing poisoning data to train more robustly.

Answer: (A)

When data poisoning exists, the lasting way to remove its influence is to unlearn the harmful patterns by retraining the model on clean, verified data. By selectively retraining with trustworthy data, the model updates its parameters based on correct signals, diminishing or replacing the erroneous associations learned from poisoned samples. This is a true long-term strategy because it directly reduces the poisoned signal in the model’s learned representations, helping restore correct behavior for future inputs. Replacing the dataset with random data would wipe out real structure and hurt performance, archiving poisoned data while keeping the same model lets the bad patterns persist, and increasing the amount of poisoned data tends to reinforce the attack rather than neutralize it.