What is the goal of long-term selective retraining in AI?

• A. To force the model to 'forget' poisoned data.
• B. To memorize all poisoned samples.
• C. To reduce model accuracy on clean data.
• D. To eliminate all training data.

Answer: (A)

The main idea here is to make the model unlearn the harmful patterns introduced by poisoned data and keep its useful knowledge intact over time. Through long-term selective retraining, you continue updating the model but focus on diminishing the influence of the poisoned samples, often by training on clean data and applying strategies that limit changes to well-functioning parts of the model. The aim is for the model to stop responding to the poisoned triggers, effectively removing the backdoor effect while preserving accuracy on legitimate inputs. This approach targets erasing those malicious associations rather than erasing all training data or harming overall performance.