Deep neural networks are sensitive to input perturbations, such as lighting changes, undermining reliability in real-world applications. This paper introduces interval-analysis enhanced MobileNet, which is a robust image classification framework that integrates interval bound propagation into the lightweight MobileNet architecture. By modeling intensity variations as bounded intervals, the method computes guaranteed output bounds that ensure prediction stability across all perturbed inputs within the defined range. The approach reveals high accuracy without any problem in computational efficiency, making it suitable for edge deployment. The network has been trained using a hybrid loss that jointly optimizes standard accuracy and robustness margins under interval uncertainty. The experiments on CIFAR-10 and ImageNet-100 demonstrate extreme performance improvements across all intensity perturbations based on natural accuracy, far exceeding adversarial training, data augmentation, and randomized smoothing. The results show that IA-MobileNet achieves a high level of accuracy while also providing formal robustness. This bridges the gap between efficiency and reliability in vision systems that perform under unpredictable environmental conditions.
Navid Razmjooy (Thu,) studied this question.