Luminance Correction based Countermeasure against Adversarial Examples

Recently, many kinds of deep learning-based image classification are introduced in various fields, such as face recognition for security authentication and autonomous driving, in order to improve their accuracy. However, previous studies have pointed out the vulnerability of them against Adversarial Examples (AE) which are able to cause misclassification intentionally. The AE utilize micro-perturbations which cannot be detected by the naked eye for causing misclassification. For practical use of the deep learning-based image classification, countermeasure against AE is a mandatory requirement. Therefore, this paper proposes countermeasures of the deep learning-based image classification against AE. The proposed method introduces luminance correction to eliminate the effect of perturbations on AE. After the proposed luminance correction, the deep learning-based image classification classifies images which are contained perturbations. Experiments prove the validity of the proposed method.