Anti-backdoor learning is a defensive machine learning paradigm designed to train secure and uncompromised neural networks directly from poisoned or untrusted datasets. Rather than relying on post-training methods that detect or repair models after they have already been compromised, this approach intervenes during the training process itself to prevent backdoor triggers from taking hold. It typically leverages differences in learning dynamics between clean and malicious data—such as the tendency of models to fit backdoor shortcuts much faster than natural task features—to identify suspected poisoned examples and suppress or unlearn the associations between malicious trigger patterns and target labels. As a result, the model retains high accuracy on standard inputs while remaining robust against trigger-based manipulation at test time.
