Adversarial Training on Limited-Resource Devices Through Asymmetric Disturbances

In the domain of adversarial training, establishing machine learning model robustness remains a paramount challenge, especially in environments constrained by resources such as mobile and IoT devices. This paper introduces a transformative framework, employing asymmetric perturbations to ensure a bounded loss and augmented computational efficiency. Our approach stands on a solid theoretical foundation and undergoes comprehensive performance evaluations. Across various metrics, we validate the efficacy of our methodology, showcasing its readiness for broad implementation in resource-limited settings.