XAI-Driven Resilient Image Classification in the Presence of Adversarial Perturbations

Deep learning (DL) architectures, although being employed in widespread applications, often raise concerns about their trustworthiness due to their opacity in their decision-making processes. Explainable AI (XAI) emerges as a promising solution to mitigate these concerns by providing interpretable rationales for DL network outputs. In domains where risk tolerance is minimal, ensuring trustworthy predictions is essential. This study introduces expmax, a new classifier rooted in XAI principles, designed for multiclass classification problems using convolutional neural network (CNN) architectures. The key strength of expmax, compared to the conventional softmax, lies in its ability to evaluate the model's focus on salient features of targets rather than being distracted by unrelated patterns from the background. This characteristic allows expmax for increased resilience, especially in scenarios with adversarial samples, where conventional classifiers may fail to correctly recognise the target class. The methodology behind expmax is based on fitting a regressor with features that are extracted from the training dataset using the SHapley Additive exPlanations (SHAP) algorithm, along with a target mask area detection algorithm. By using the SHAP-based extracted features, expmax reduces vulnerabilities to perturbations introduced by adversarial inputs. The method is validated on the MTARSI dataset for aircraft recognition in remote sensing images.