Abstract: Artificial Intelligence (AI) has become integral to modern robotics and intelligent systems, but its increasing deployment in critical domains raises significant safety and explainability challenges. In this talk, we explore the limitations of current AI models, including adversarial vulnerabilities, distribution shifts, and failures in safety-critical contexts. We introduce SafeML, a statistical distance–based framework for monitoring concept drift, distribution shifts, and domain adaptation, offering robust safety assurance and human-in-the-loop oversight. Additionally, we present SMILE (Statistical Model-agnostic Interpretability with Local Explanations), a novel model-agnostic explainability approach that enhances robustness against adversarial attacks compared to LIME and SHAP. Case studies across various domains, including intrusion detection, traffic sign recognition, medical diagnosis, and robotics, demonstrate the practical application and reproducibility of these methods. Together, SafeML and SMILE provide a pathway toward safer, more interpretable, and trustworthy AI for critical systems.
[Disponível]MSSiS Copyright 2025. Todos os direitos reservados. Design by Free html Templates
