Attribution Robustness

Attribution robustness in machine learning focuses on ensuring that explanations of model predictions remain reliable even when the input data is slightly altered. Current research investigates the vulnerability of various attribution methods, including gradient-based and removal-based approaches, across different model architectures like transformers and convolutional neural networks, and explores techniques like adversarial training and smoothing to enhance robustness. This field is crucial for building trust in AI systems, particularly in high-stakes applications, by guaranteeing the fidelity and reliability of model explanations.