DCAL: A New Method for Defending Against Adversarial Examples

In recent years, deep learning has shown excellent performance in the field of computer vision. Nevertheless, researchers have found that the deep learning system does not have good robustness. Adding a insignificant amount of undetectable interference to the input of the deep learning system can lead to deep learning models fail, and these examples that make the model fail are called adversarial examples by researchers. The existence of adversarial examples will hinder the application and popularization of artificial intelligence-based deep learning systems. Therefore, we propose a denoising convolutional autoencoder incorporated with label knowledge (DCAL), a new method for defending against adversarial examples. The principle of which is DCAL as a pre-processing module before image classification, the image to be classified is denoised and reconstructed to obtain a innovative image, which is then sent to the classifier for classification. If we let the innovative image obtained by the adversarial examples through DCAL can make the classifier classify correctly, we will achieve the role of defending against the adversarial examples. The experimental results on two benchmark datasets including MNIST, CIFAR-10. Our experimental principally resisting the white-box attacks. The experimental results show that the proposed DCAL is superior to state-of-the-art defense methods in a white-box setting.