Backdoor attacks are a growing concern in machine learning, where adversaries slip poisoned samples into training data. These samples carry triggers that can flip a model's predictions on specific inputs while leaving its overall behavior intact, according to a new arXiv paper.
The paper introduces SemPurify, a semantics-aware data purification approach designed for black-box backdoor defense. Unlike white-box methods that require access to model internals, SemPurify operates in a black-box setting, making it more practical for real-world deployments.
By focusing on semantics, the method aims to remove or neutralize triggers without degrading the model's performance on clean data. The abstract does not provide experimental details, so the effectiveness of SemPurify relative to existing defenses remains to be seen.