| Abstract [eng] |
After training an artificial neural network, it is often necessary to interpret it as a black box. However, there are several reasons why it is important to be able to explain and interpret the results produced by artificial neural networks. Understanding why the model made a certain decision increases the reliability of systems, opens up possibilities for improving neural network models, and provides a new source of learning. LIME, SHAP, and Grad-CAM are explainable artificial intelligence methods that can help bring transparency to very complex and unclear models, uncover the reasoning behind the model’s decisions, and detect adversarial attacks in which an image is corrupted with noise that is invisible to the human eye but affects how the neural network model classifies it. By applying explainable AI methods (LIME, Grad-CAM, SHAP), it is possible to visually evaluate the features that influenced the classification result. Additionally, based on the obtained SHAP numerical pixel importance values, it is possible to determine the likelihood that an image has been affected by a noise-based attack. It has been observed that features of a noise-free image are more significant and concentrated on specific visible elements in the image, whereas explanations generated for an attacked image are less focused on a few specific features, their importance is more randomly dispersed across the entire image. |
