Self-Guided Integrated Gradient Method for Attribution

Explaining the decisions of deep neural networks is essential for building trust in AI systems, particularly when deployed in sensitive domains such as healthcare, security, and autonomous transport. Path-based attribution methods such as Integrated Gradients provide local explanations by integrating model gradients along a path from a baseline to the input image. However, such methods require user-defined baselines and often accumulate noisy gradients from saturated regions of the prediction landscape. To address these limitations, we propose the Self-guided Integrated Gradient Method for Attribution (SIGMA), a baseline-free path-based method that stochastically explores the model’s confidence landscape to identify input features responsible for the collapse of class confidence. By following the model’s decision boundary, SIGMA produces interpretable and reliable attributions without requiring reference inputs or access to a model’s internal representation layers. Evaluations across diverse architectures, including Vision Transformers, and computer vision datasets in healthcare and security, demonstrate that SIGMA provides spatially coherent attribution maps with strong faithfulness to the model’s internal reasoning. Additionally, SIGMA generates zero-confidence variants of the input, recognisable to humans but adversarial to the model. We show that augmenting the original training dataset and retraining with these samples improves both robustness to noise and resistance to adversarial attacks.