Persist BD

AI model backdoors persist through retraining, posing security risks

Researchers have demonstrated that backdoors can be easily implanted into AI models and are difficult to remove, even after retraining. In experiments, a hidden behavior triggered by a specific comment was successfully …