Abstract
We present a generic Reinforcement Learning (RL) framework optimized forcrafting adversarial attacks on different model types spanning from ECG signalanalysis (1D), image classification (2D), and video classification (3D). Theframework focuses on identifying sensitive regions and inducingmisclassifications with minimal distortions and various distortion types. Thenovel RL method outperforms state-of-the-art methods for all threeapplications, proving its efficiency. Our RL approach produces superiorlocalization masks, enhancing interpretability for image classification and ECGanalysis models. For applications such as ECG analysis, our platform highlightscritical ECG segments for clinicians while ensuring resilience againstprevalent distortions. This comprehensive tool aims to bolster both resiliencewith adversarial training and transparency across varied applications and datatypes.