Deep neural networks have achieved state-of-the-art performance across a wide range of computer vision tasks, yet their deployment in real-world and safety-critical domains remains limited by two major challenges: vulnerability to adversarial perturbations and high computational cost.
