• Journal of Korea Multimedia Society
• /
• v.25 no.6
• /
• pp.878-885
• /
• 2022

Research in the field of computer vision based on deep learning is being actively conducted. However, deep learning-based models have vulnerabilities in adversarial attacks that increase the model's misclassification rate by applying adversarial perturbation. In particular, in the case of FGSM, it is recognized as one of the effective attack methods because it is simple, fast and has a considerable attack success rate. Meanwhile, as one of the efforts to visualize deep learning models, Grad-CAM enables visual explanation of convolutional neural networks. In this paper, I propose a method to generate adversarial examples with high attack success rate by applying Grad-CAM to FGSM. The method chooses fixels, which are closely related to labels, by using Grad-CAM and add perturbations to the fixels intensively. The proposed method has a higher success rate than the FGSM model in the same perturbation for both targeted and untargeted examples. In addition, unlike FGSM, it has the advantage that the distribution of noise is not uniform, and when the success rate is increased by repeatedly applying noise, the attack is successful with fewer iterations.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.15 no.1
• /
• pp.27-35
• /
• 2022

In this paper we propose white noise adding method to prevent missclassification of deep learning system by adversarial attacks. The proposed method is that adding white noise to input image that is benign or adversarial example. The experimental results are showing that the proposed method is robustness to 3 adversarial attacks such as FGSM attack, BIN attack and CW attack. The recognition accuracies of Resnet model with 18, 34, 50 and 101 layers are enhanced when white noise is added to test data set while it does not affect to classification of benign test dataset. The proposed model is applicable to defense to adversarial attacks and replace to time- consuming and high expensive defense method against adversarial attacks such as adversarial training method and deep learning replacing method.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.29 no.1
• /
• pp.117-125
• /
• 2019

Deep Learning, which is being used in various fields recently, is being threatened with Adversarial Attack. In this paper, we experimentally verify that the classification accuracy is lowered by adversarial samples generated by malicious attackers in image classification models. We used MNIST dataset and measured the detection accuracy by injecting adversarial samples into the Autoencoder classification model and the CNN (Convolution neural network) classification model, which are created using the Tensorflow library and the Pytorch library. Adversarial samples were generated by transforming MNIST test dataset with JSMA(Jacobian-based Saliency Map Attack) and FGSM(Fast Gradient Sign Method). When injected into the classification model, detection accuracy decreased by at least 21.82% up to 39.08%.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.31 no.6
• /
• pp.1215-1225
• /
• 2021

Although Artificial Intelligence (AI) techniques have shown impressive performance in various fields, they are vulnerable to adversarial examples which induce misclassification by adding human-imperceptible perturbations to the input. Previous studies to defend the adversarial examples can be classified into three categories: (1) model retraining methods; (2) input transformation methods; and (3) adversarial examples detection methods. However, even though the defense methods against adversarial examples have constantly been proposed, there is no research to classify the type of adversarial attack. In this paper, we proposed an adversarial attack family classification method based on dimensionality reduction and clustering. Specifically, after extracting adversarial perturbation from adversarial example, we performed Linear Discriminant Analysis (LDA) to reduce the dimensionality of adversarial perturbation and performed K-means algorithm to classify the type of adversarial attack family. From the experimental results using MNIST dataset and CIFAR-10 dataset, we show that the proposed method can efficiently classify five tyeps of adversarial attack(FGSM, BIM, PGD, DeepFool, C&W). We also show that the proposed method provides good classification performance even in a situation where the legitimate input to the adversarial example is unknown.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.6
• /
• pp.907-917
• /
• 2023

Recently, a self-driving car have applied deep learning technology to advanced driver assistance system can provide convenience to drivers, but it is shown deep that learning technology is vulnerable to adversarial evasion attacks. In this paper, we performed five adversarial evasion attacks, including MI-FGSM(Momentum Iterative-Fast Gradient Sign Method), targeting the object detection algorithm YOLOv5 (You Only Look Once), and measured the object detection performance in terms of mAP(mean Average Precision). In particular, we present a method applying morphology operations for YOLO to detect objects normally by removing noise and extracting boundary. As a result of analyzing its performance through experiments, when an adversarial attack was performed, YOLO's mAP dropped by at least 7.9%. The YOLO applied our proposed method can detect objects up to 87.3% of mAP performance.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.05a
• /
• pp.404-407
• /
• 2019

최근 많은 영역에 딥러닝이 활용되고 있다. 특히 CNN과 같은 아키텍처는 얼굴인식과 같은 이미지 분류 분야에서 활용된다. 이러한 딥러닝 기술을 완전한 기술로서 활용할 수 있는지에 대한 연구가 이뤄져왔다. 관련 연구로 PGD(Projected Gradient Descent) 공격이 존재한다. 해당 공격을 이용하여 원본 이미지에 노이즈를 더해주게 되면, 수정된 이미지는 전혀 다른 클래스로 분류되게 된다. 본 연구에서 기존의 FGSM(Fast gradient sign method) 공격기법에 Triplet loss를 활용한 Adversarial 공격 모델을 제안 및 구현하였다. 제안된 공격 모델은 간단한 시나리오를 기반으로 검증하였고 해당 결과를 분석하였다.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.4
• /
• pp.621-631
• /
• 2023

Adversarial training improves the robustness of deep neural networks for adversarial examples. However, the previous adversarial training method focuses only on the adversarial loss function, ignoring that even a small perturbation of the input layer causes a significant change in the hidden layer features. Consequently, the accuracy of a defended model is reduced for various untrained situations such as clean samples or other attack techniques. Therefore, an architectural perspective is necessary to improve feature representation power to solve this problem. In this paper, we apply an attention module that generates an attention map of an input image to a general model and performs PGD adversarial training upon the augmented model. In our experiments on the CIFAR-10 dataset, the attention augmented model showed higher accuracy than the general model regardless of the network structure. In particular, the robust accuracy of our approach was consistently higher for various attacks such as PGD, FGSM, and BIM and more powerful adversaries. By visualizing the attention map, we further confirmed that the attention module extracts features of the correct class even for adversarial examples.

• Journal of the Institute of Convergence Signal Processing
• /
• v.23 no.4
• /
• pp.194-199
• /
• 2022

Preprocessing for high-quality data is required for high accuracy and usability in various and complex image data-based industries. However, when a contaminated hostile example that combines noise with existing image or video data is introduced, which can pose a great risk to the company, it is necessary to restore the previous damage to ensure the company's reliability, security, and complete results. As a countermeasure for this, restoration was previously performed using Defense-GAN, but there were disadvantages such as long learning time and low quality of the restoration. In order to improve this, this paper proposes a method using adversarial examples created through FGSM according to image segmentation in addition to using the VQ-VAE model. First, the generated examples are classified as a general classifier. Next, the unsegmented data is put into the pre-trained VQ-VAE model, restored, and then classified with a classifier. Finally, the data divided into quadrants is put into the 4-split-VQ-VAE model, the reconstructed fragments are combined, and then put into the classifier. Finally, after comparing the restored results and accuracy, the performance is analyzed according to the order of combining the two models according to whether or not they are split.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.12 no.6
• /
• pp.629-635
• /
• 2019

In Deep Learning models derivative is implemented by error back-propagation which enables the model to learn the error and update parameters. It can find the global (or local) optimal points of parameters even in the complex models taking advantage of a huge improvement in computing power. However, deliberately generated data points can 'fool' models and degrade the performance such as prediction accuracy. Not only these adversarial examples reduce the performance but also these examples are not easily detectable with human's eyes. In this work, we propose the method to detect adversarial datasets with random noise addition. We exploit the fact that when random noise is added, prediction accuracy of non-adversarial dataset remains almost unchanged, but that of adversarial dataset changes. We set attack methods (FGSM, Saliency Map) and noise level (0-19 with max pixel value 255) as independent variables and difference of prediction accuracy when noise was added as dependent variable in a simulation experiment. We have succeeded in extracting the threshold that separates non-adversarial and adversarial dataset. We detected the adversarial dataset using this threshold.