Deep Neural Networks (DNNs) have shown outstanding performance on a variety of tasks. However, their vulnerability to adversarial examples has become a significant concern in recent years. Of particular concern are imperceptible adversarial examples, which present a latent threat to safety-critical applications. In this work, we introduce an Attention-Gated Denoising Autoencoder model (AGDAE) designed to effectively detect imperceptible adversarial attacks. Our model utilizes innovative Attention-Gated Convolutional Blocks to enhance feature extraction and selection, thereby improving the model’s sensitivity to minute perturbations. By using an autoencoder model, our network is able to remove adversarial noise by reconstructing input images. Additionally, we combine a target-oriented detection framework to verify label consistency between reconstructed and original images, further enhancing adversarial example detection. Experiments on datasets such as CIFAR-10, and ImageNet validate the effectiveness of AGDAE in detecting imperceptible adversarial examples, maintaining a high detection accuracy even against highly sophisticated adversarial attacks.