Detecting and Defending Adversarial Attacks on Deep Learning Models Using Convolutional Autoencoders and Block-Switching ResNet

Deep learning, particularly convolutional neural networks (CNNs), has proven valuable for brain tumor classification, aiding diagnostic and therapeutic decisions in medical imaging. Despite their accuracy, these models are vulnerable to adversarial attacks, compromising their reliability in clinical settings. In this research, we utilized a VGG16-based CNN model to classify brain tumors, achieving 96% accuracy on clean magnetic resonance imaging (MRI) data. To assess robustness, we exposed the model to Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks, which reduced accuracy to 32% and 13%, respectively. We then applied a multi-layered defense strategy, including adversarial training with FGSM and PGD examples and feature squeezing techniques such as bit-depth reduction and Gaussian blurring. This approach improved model resilience, achieving 54% accuracy on FGSM and 47% on PGD adversarial examples. Our results highlight the importance of proactive defense strategies for maintaining the reliability of AI in medical imaging under adversarial conditions.

With the widespread use of deep learning models in various applications. People are gradually realizing the vulnerability of these models to adversarial attacks. Adversarial training is an effective strategy to defend against adversarial attacks. Based on the advantages and disadvantages of the current mainstream Fast Gradient Sign Method (FGSM) adversarial training and Projected Gradient Descent (PGD) adversarial training, this paper proposes a hybrid adversarial training that integrates FGSM and PGD methods and uses the ResNet-18 model and SVHN dataset for testing. Experimental results show that hybrid adversarial training can effectively reduce training time. Its accuracy on the original data set is higher than that of PGD adversarial training, which is improved by about 2%. The performance when facing FGSM attacks is almost the same as that of single FGSM adversarial training. The performance when facing PGD attacks decreases more significantly, which is about 2% to 3% lower than that of PGD adversarial training. This study not only helps to understand the robustness of hybrid adversarial training to models facing adversarial attacks but also helps in studying new adversarial training strategies.

In this paper, a novel black box adversarial computer vision attack is proposed. The introduced attack is based on removing from images some components described by their Tchebichef discrete orthogonal moments, rather than to perturb them. The contribution of this work is focused on the addition of one more clue, supporting the critical hypothesis that computer vision systems fail because they support their decisions not only in robust features but also in others non-robust ones. In this, context non-robust image features described in terms of Tchebichef moments are excluded from the original images and the approximated reconstructed versions of them are used as adversarial examples in order to attack some popular deep learning models. The experiments justify the effectiveness of the proposed adversarial attack in terms of imperceptibility and recognition error rate of the deep learning classifiers. It is worth noting that the top-1 accuracy of the attacked models was degraded by a factor between 9.48%-70.89% for adversarial images of 65dB to 57dB PSNR values. The corresponding degradation of the top-5 models’ accuracy was between 6.9% and 55.14% for the same quality images. Moreover, the proposed attack seems to have more strength than the Fast Gradient Sign Method (FGSM) attacking method traditionally applying in most cases. These results reveal that the proposed attack is able to exploit the vulnerability of the deep learning models’ towards degrading their generalization abilities. In this paper, a novel black box adversarial computer vision attack is proposed. The introduced attack is based on removing from images some components described by their Tchebichef discrete orthogonal moments, rather than to perturb them. The contribution of this work is focused on the addition of one more clue, supporting the critical hypothesis that computer vision systems fail because they support their decisions not only in robust features but also in others non-robust ones. In this, context non-robust image features described in terms of Tchebichef moments are excluded from the original images and the approximated reconstructed versions of them are used as adversarial examples in order to attack some popular deep learning models. The experiments justify the effectiveness of the proposed adversarial attack in terms of imperceptibility and recognition error rate of the deep learning classifiers. It is worth noting that the top-1 accuracy of the attacked models was degraded by a factor between 9.48%-70.89% for adversarial images of 65dB to 57dB PSNR values. The corresponding degradation of the top-5 models’ accuracy was between 6.9% and 55.14% for the same quality images. Moreover, the proposed attack seems to have more strength than the Fast Gradient Sign Method (FGSM) attacking method traditionally applying in most cases. These results reveal that the proposed attack is able to exploit the vulnerability of the deep learning models’ towards degrading their generalization abilities.

Untargeted white-box adversarial attack with heuristic defence methods in real-time deep learning based network intrusion detection system

Deep learning has revolutionized image recognition. One significant obstacle still remains, the vulnerability of these models to adversarial attacks. These attacks manipulate images with subtle changes that cause CNN misclassification. While methods, such as adversarial training, have been proposed to defend against adversarial attacks, they incur additional training costs, either extra input samples or auxiliary models. In this work, we propose an efficient approach to deploying robust models that utilize image transformations to remove adversarial noises. We investigate the performance of simple transformations and report several effective ones, including Affine blur, Gaussian blur, Median blur, and Bilateral blur against various adversarial attack methods, such as Fast Gradient Sign Method (FGSM), Randomized + FGSM (RFGSM), and Projected Gradient Descent (PGD). We apply these image transformation techniques to the widely used ImageNet dataset, and experimental results demonstrate the potential of image transformation methods as a strong defense against adversarial attacks in deep learning-based image classification systems, especially when combined with cutting-edge neural network architectures such as ResNet50 and DenseNet121. Our comprehensive results show that these transformations can significantly improve the robustness of CNN models against adversarial attacks on ImageNet, achieving a recovery rate of up to 85% to 90% without incurring extra resource costs.

The design of a security scheme for beamforming prediction is critical for next-generation wireless networks (5G, 6G, and beyond). However, there is no consensus about protecting beamforming prediction using deep learning algorithms in these networks. This paper presents the security vulnerabilities in deep learning for beamforming prediction using deep neural networks in 6G wireless networks, which treats the beamforming prediction as a multi-output regression problem. It is indicated that the initial DNN model is vulnerable to adversarial attacks, such as Fast Gradient Sign Method , Basic Iterative Method , Projected Gradient Descent , and Momentum Iterative Method , because the initial DNN model is sensitive to the perturbations of the adversarial samples of the training data. This study offers two mitigation methods, such as adversarial training and defensive distillation, for adversarial attacks against artificial intelligence-based models used in the millimeter-wave (mmWave) beamforming prediction. Furthermore, the proposed scheme can be used in situations where the data are corrupted due to the adversarial examples in the training data. Experimental results show that the proposed methods defend the DNN models against adversarial attacks in next-generation wireless networks.

ZeroGrad: Costless conscious remedies for catastrophic overfitting in the FGSM adversarial training

In this paper, we propose a new method for detecting adversarial attacks on deep neural networks. Our algorithm is based on the intuition that attacking input images results in different displacement vectors for clean and adversarial classes. For example, if the input image is an adversarial example, the re-attacking process results in a displacement vector with a short length in the feature space, but this displacement is considerable for clean images. We train our detector based on these displacement vectors. The experimental results show that compared to the current learning-based adversarial detection methods, the proposed system is capable of detecting the adversarial examples using a far simpler network. In addition, the proposed method is independent of the attack type, and is able to detect even novel attacks. It is also revealed that the proposed system learns the discrimination function even using a small amount of training data without any hyper-parameter tuning. We obtain remarkable results in detecting adversarial examples which are placed near and far from the decision boundary, improving state-of-the-art in detecting 2-norm Carlini and Wagner attack (L2-C&W) and ∞-norm Projected Gradient Descent attack (L∞-PGD), where just Fast Gradient Sign Method (FGSM) is used for training the system.

This chapter introduces the concept of adversarial attacks on image classification models built on convolutional neural networks (CNN). CNNs are very popular deep-learning models which are used in image classification tasks. However, very powerful and pre-trained CNN models working very accurately on image datasets for image classification tasks may perform disastrously when the networks are under adversarial attacks. In this work, two very well-known adversarial attacks are discussed and their impact on the performance of image classifiers is analyzed. These two adversarial attacks are the fast gradient sign method (FGSM) and adversarial patch attack. These attacks are launched on three powerful pre-trained image classifier architectures, ResNet-34, GoogleNet, and DenseNet-161. The classification accuracy of the models in the absence and presence of the two attacks are computed on images from the publicly accessible ImageNet dataset. The results are analyzed to evaluate the impact of the attacks on the image classification task.

Deep learning architectures have emerged as powerful function approximators in a broad spectrum of complex representation learning tasks, such as, computer vision, natural language processing and collaborative filtering. These architectures bear a high potential to learn the intrinsic structure of data and extract valuable insights. Despite the surge in the development of state-of-the-art intelligent systems using the deep neural networks (DNNs), these systems have found to be vulnerable to adversarial examples produced by adding a small-magnitude of perturbations. Such adversarial examples are adept at misleading the DNN classifiers. In the past, different attack strategies have been proposed to produce adversarial examples in the digital, physical, and transform domain, but the likelihood to generate perceptually realistic adversarial examples require more research efforts. In this paper, we present a novel approach to produce adversarial examples by combining the single-shot fast gradient sign method (FGSM) and spatial, as well as, transform domain image processing techniques. The resulted perturbations neutralize the impact of low-intensity based regions, thus, instilling the noise only in the selective high-intensity regions of the input image. While combining the customized perturbation with one-step FGSM perturbation in an un-targeted black-box attack scenario, the proposed approach successfully fools state-of-the-art DNN classifiers with 99% adversarial examples being misclassified on the ImageNet validation dataset.

Deep neural networks (DNN’s) have achieved state-of-the-art (SOTA) for many image-related tasks like classification, object detection, etc. The authors have shown that DNN’s can be easily fooled by adversarial examples that can be generated by adding small distortions or perturbations to the images. Several attack methods to add these perturbations in images have been proposed like the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), etc. to create distortions in images that cause the DNN’s to misclassify instances. It is essential to come up with strategies to identify such fake images. Keeping this in mind, we propose an end-to-end pipeline capable of detecting adversarially attacked images, i.e., images with perturbations. For a given image, we generate images with manipulation masks using Mantra-Net and pass them through a trained binary classifier that distinguishes between perturbed and original images. Experiments on the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2012 dataset have shown that our pipeline achieves an accuracy of 0.986 on the task of detecting adversarially crafted images.

Although deep neural networks (DNNs) have achieved high performance across various applications, they are often deceived by adversarial examples generated by adding small perturbations. To combat adversarial attacks, many detection methods have been proposed, including feature squeezing and trapdoor. However, these methods rely on the output of DNNs or involve training a separate network to detect adversarial examples, which leads to high computational costs and low efficiency. In this study, we propose a simple and effective approach called the entropy-based detector (EBD) to protect DNNs from various adversarial attacks. EBD detects adversarial examples by comparing the difference in entropy between the input sample before and after bit depth reduction. We show that EBD can detect over 98% of the adversarial examples generated by attacks using fast-gradient sign method, basic iterative method, momentum iterative method, DeepFool and CW attacks when the false positive rate is 2.5% for CIFAR-10 and ImageNet datasets.

Although fast adversarial training has demonstrated both robustness and efficiency, the problem of "catastrophic overfitting" has been observed. This is a phenomenon in which, during single-step adversarial training, the robust accuracy against projected gradient descent (PGD) suddenly decreases to 0% after a few epochs, whereas the robust accuracy against fast gradient sign method (FGSM) increases to 100%. In this paper, we demonstrate that catastrophic overfitting is very closely related to the characteristic of single-step adversarial training which uses only adversarial examples with the maximum perturbation, and not all adversarial examples in the adversarial direction, which leads to decision boundary distortion and a highly curved loss surface. Based on this observation, we propose a simple method that not only prevents catastrophic overfitting, but also overrides the belief that it is difficult to prevent multi-step adversarial attacks with single-step adversarial training.

The successful emergence of deep learning (DL) in wireless system applications has raised concerns about new security-related challenges. One such security challenge is adversarial attacks. Although there has been much work demonstrating the susceptibility of DL-based classification tasks to adversarial attacks, regression-based problems in the context of a wireless system have not been studied so far from an attack perspective. The aim of this paper is twofold: (i) we consider a regression problem in a wireless setting and show that adversarial attacks can break the DL-based approach and (ii) we analyze the effectiveness of adversarial training as a defensive technique in adversarial settings and show that the robustness of DL-based wireless system against attacks improves significantly. Specifically, the wireless application considered in this paper is the DL-based power allocation in the downlink of a multicell massive multi-input-multi-output system, where the goal of the attack is to yield an infeasible solution by the DL model. We extend the gradient-based adversarial attacks: fast gradient sign method (FGSM), momentum iterative FGSM, and projected gradient descent method to analyze the susceptibility of the considered wireless application with and without adversarial training. We analyze the deep neural network (DNN) models performance against these attacks, where the adversarial perturbations are crafted using both the white-box and black-box attacks.

To improve the robustness of intrusion detection systems constructed using deep learning models, a method based on an auxiliary adversarial training WGAN (AuxAtWGAN) is proposed from the defender’s perspective. First, one-dimensional traffic data are downscaled and processed into two-dimensional image data via a stacked autoencoder (SAE), and mixed adversarial samples are generated using the fast gradient sign method (FGSM), Projected Gradient Descent (PGD) and Carlini and Wagner (C&W) adversarial attacks. Second, the improved WGAN with an integrated perceptual network module is trained with mixed training samples composed of mixed adversarial samples and normal samples. Finally, the adversary-trained AuxAtWGAN model is attached to the original model for adversary sample detection, and the detected adversary samples are removed and input into the original model to improve the robustness of the original model. The average attack success rate of the original convolutional neural network (CNN) model against multiple adversarial samples is 75.17%, and after using AuxAtWGAN, the average attack success rate of the adversarial attacks decreases to 27.56%; moreover, the detection accuracy of the original CNN model against normal samples is still 93.57%. The experiment proves that AuxAtWGAN improves the robustness of the original model. In addition, validation experiments are conducted by attaching the AuxAtWGAN model to the Long Short-Term Memory Network (LSTM) and Residual Network34 (ResNet) models, which prove that the proposed method has high generalization performance.