HARVEY: New Method Removes Neural Network Backdoors Effectively

A new research paper introduces HARVEY, an innovative technique designed to combat neural network backdoors, which are often introduced through data poisoning during training. Unlike previous methods that focus on identifying benign samples, HARVEY takes a different approach by learning an oracle specifically for poisonous samples. This distinction is crucial because learning a backdoored reference model is considerably easier than learning one based on benign data. This ease of identifying poisonous samples allows HARVEY to achieve much higher accuracy in detecting them compared to related work. The method has been rigorously evaluated across various attack types, datasets, and network architectures, demonstrating its superior effectiveness. HARVEY consistently reduces the attack success rate to a minimum while maintaining the model's natural accuracy, offering a robust solution for enhancing the security and trustworthiness of AI systems.