Philipp Benz

Philipp Benz, Chaoning Zhang, Adil Karjauv et al.

Despite their overwhelming success on a wide range of applications, convolutional neural networks (CNNs) are widely recognized to be vulnerable to adversarial examples. This intriguing phenomenon led to a competition between adversarial attacks and defense techniques. So far, adversarial training is the most widely used method for defending against adversarial attacks. It has also been extended to defend against universal adversarial perturbations (UAPs). The SOTA universal adversarial training (UAT) method optimizes a single perturbation for all training samples in the mini-batch. In this work, we find that a UAP does not attack all classes equally. Inspired by this observation, we identify it as the source of the model having unbalanced robustness. To this end, we improve the SOTA UAT by proposing to utilize class-wise UAPs during adversarial training. On multiple benchmark datasets, our class-wise UAT leads superior performance for both clean accuracy and adversarial robustness against universal attack.

Chaoning Zhang, Philipp Benz, Chenguo Lin et al.

The intriguing phenomenon of adversarial examples has attracted significant attention in machine learning and what might be more surprising to the community is the existence of universal adversarial perturbations (UAPs), i.e. a single perturbation to fool the target DNN for most images. With the focus on UAP against deep classifiers, this survey summarizes the recent progress on universal adversarial attacks, discussing the challenges from both the attack and defense sides, as well as the reason for the existence of UAP. We aim to extend this work as a dynamic survey that will regularly update its content to follow new works regarding UAP or universal attack in a wide range of domains, such as image, audio, video, text, etc. Relevant updates will be discussed at: https://bit.ly/2SbQlLG. We welcome authors of future works in this field to contact us for including your new finding.

Chaoning Zhang, Philipp Benz, Adil Karjauv et al.

The booming interest in adversarial attacks stems from a misalignment between human vision and a deep neural network (DNN), i.e. a human imperceptible perturbation fools the DNN. Moreover, a single perturbation, often called universal adversarial perturbation (UAP), can be generated to fool the DNN for most images. A similar misalignment phenomenon has recently also been observed in the deep steganography task, where a decoder network can retrieve a secret image back from a slightly perturbed cover image. We attempt explaining the success of both in a unified manner from the Fourier perspective. We perform task-specific and joint analysis and reveal that (a) frequency is a key factor that influences their performance based on the proposed entropy metric for quantifying the frequency distribution; (b) their success can be attributed to a DNN being highly sensitive to high-frequency content. We also perform feature layer analysis for providing deep insight on model generalization and robustness. Additionally, we propose two new variants of universal perturbations: (1) Universal Secret Adversarial Perturbation (USAP) that simultaneously achieves attack and hiding; (2) high-pass UAP (HP-UAP) that is less visible to the human eye.

Chaoning Zhang, Adil Karjauv, Philipp Benz et al.

Data hiding is one widely used approach for protecting authentication and ownership. Most multimedia content like images and videos are transmitted or saved in the compressed form. This kind of lossy compression, such as JPEG, can destroy the hidden data, which raises the need of robust data hiding. It is still an open challenge to achieve the goal of data hiding that can be against these compressions. Recently, deep learning has shown large success in data hiding, while non-differentiability of JPEG makes it challenging to train a deep pipeline for improving robustness against lossy compression. The existing SOTA approaches replace the non-differentiable parts with differentiable modules that perform similar operations. Multiple limitations exist: (a) large engineering effort; (b) requiring a white-box knowledge of compression attacks; (c) only works for simple compression like JPEG. In this work, we propose a simple yet effective approach to address all the above limitations at once. Beyond JPEG, our approach has been shown to improve robustness against various image and video lossy compression algorithms.

Philipp Benz, Chaoning Zhang, Adil Karjauv et al.

Convolutional neural networks (CNNs) have made significant advancement, however, they are widely known to be vulnerable to adversarial attacks. Adversarial training is the most widely used technique for improving adversarial robustness to strong white-box attacks. Prior works have been evaluating and improving the model average robustness without class-wise evaluation. The average evaluation alone might provide a false sense of robustness. For example, the attacker can focus on attacking the vulnerable class, which can be dangerous, especially, when the vulnerable class is a critical one, such as "human" in autonomous driving. We propose an empirical study on the class-wise accuracy and robustness of adversarially trained models. We find that there exists inter-class discrepancy for accuracy and robustness even when the training dataset has an equal number of samples for each class. For example, in CIFAR10, "cat" is much more vulnerable than other classes. Moreover, this inter-class discrepancy also exists for normally trained models, while adversarial training tends to further increase the discrepancy. Our work aims to investigate the following questions: (a) is the phenomenon of inter-class discrepancy universal regardless of datasets, model architectures and optimization hyper-parameters? (b) If so, what can be possible explanations for the inter-class discrepancy? (c) Can the techniques proposed in the long tail classification be readily extended to adversarial training for addressing the inter-class discrepancy?

Chaoning Zhang, Philipp Benz, Dawit Mureja Argaw et al.

ResNet or DenseNet? Nowadays, most deep learning based approaches are implemented with seminal backbone networks, among them the two arguably most famous ones are ResNet and DenseNet. Despite their competitive performance and overwhelming popularity, inherent drawbacks exist for both of them. For ResNet, the identity shortcut that stabilizes training also limits its representation capacity, while DenseNet has a higher capacity with multi-layer feature concatenation. However, the dense concatenation causes a new problem of requiring high GPU memory and more training time. Partially due to this, it is not a trivial choice between ResNet and DenseNet. This paper provides a unified perspective of dense summation to analyze them, which facilitates a better understanding of their core difference. We further propose dense weighted normalized shortcuts as a solution to the dilemma between them. Our proposed dense shortcut inherits the design philosophy of simple design in ResNet and DenseNet. On several benchmark datasets, the experimental results show that the proposed DSNet achieves significantly better results than ResNet, and achieves comparable performance as DenseNet but requiring fewer computation resources.

Philipp Benz, Chaoning Zhang, Adil Karjauv et al.

The performance of DNNs trained on clean images has been shown to decrease when the test images have common corruptions. In this work, we interpret corruption robustness as a domain shift and propose to rectify batch normalization (BN) statistics for improving model robustness. This is motivated by perceiving the shift from the clean domain to the corruption domain as a style shift that is represented by the BN statistics. We find that simply estimating and adapting the BN statistics on a few (32 for instance) representation samples, without retraining the model, improves the corruption robustness by a large margin on several benchmark datasets with a wide range of model architectures. For example, on ImageNet-C, statistics adaptation improves the top1 accuracy of ResNet50 from 39.2% to 48.7%. Moreover, we find that this technique can further improve state-of-the-art robust models from 58.1% to 63.3%.

Philipp Benz, Chaoning Zhang, In So Kweon

Batch normalization (BN) has been widely used in modern deep neural networks (DNNs) due to improved convergence. BN is observed to increase the model accuracy while at the cost of adversarial robustness. There is an increasing interest in the ML community to understand the impact of BN on DNNs, especially related to the model robustness. This work attempts to understand the impact of BN on DNNs from a non-robust feature perspective. Straightforwardly, the improved accuracy can be attributed to the better utilization of useful features. It remains unclear whether BN mainly favors learning robust features (RFs) or non-robust features (NRFs). Our work presents empirical evidence that supports that BN shifts a model towards being more dependent on NRFs. To facilitate the analysis of such a feature robustness shift, we propose a framework for disentangling robust usefulness into robustness and usefulness. Extensive analysis under the proposed framework yields valuable insight on the DNN behavior regarding robustness, e.g. DNNs first mainly learn RFs and then NRFs. The insight that RFs transfer better than NRFs, further inspires simple techniques to strengthen transfer-based black-box attacks.

Chaoning Zhang, Philipp Benz, Tooba Imtiaz et al.

A single universal adversarial perturbation (UAP) can be added to all natural images to change most of their predicted class labels. It is of high practical relevance for an attacker to have flexible control over the targeted classes to be attacked, however, the existing UAP method attacks samples from all classes. In this work, we propose a new universal attack method to generate a single perturbation that fools a target network to misclassify only a chosen group of classes, while having limited influence on the remaining classes. Since the proposed attack generates a universal adversarial perturbation that is discriminative to targeted and non-targeted classes, we term it class discriminative universal adversarial perturbation (CD-UAP). We propose one simple yet effective algorithm framework, under which we design and compare various loss function configurations tailored for the class discriminative universal attack. The proposed approach has been evaluated with extensive experiments on various benchmark datasets. Additionally, our proposed approach achieves state-of-the-art performance for the original task of UAP attacking all classes, which demonstrates the effectiveness of our approach.

Philipp Benz, Chaoning Zhang, Tooba Imtiaz et al.

Despite their impressive performance, deep neural networks (DNNs) are widely known to be vulnerable to adversarial attacks, which makes it challenging for them to be deployed in security-sensitive applications, such as autonomous driving. Image-dependent perturbations can fool a network for one specific image, while universal adversarial perturbations are capable of fooling a network for samples from all classes without selection. We introduce a double targeted universal adversarial perturbations (DT-UAPs) to bridge the gap between the instance-discriminative image-dependent perturbations and the generic universal perturbations. This universal perturbation attacks one targeted source class to sink class, while having a limited adversarial effect on other non-targeted source classes, for avoiding raising suspicions. Targeting the source and sink class simultaneously, we term it double targeted attack (DTA). This provides an attacker with the freedom to perform precise attacks on a DNN model while raising little suspicion. We show the effectiveness of the proposed DTA algorithm on a wide range of datasets and also demonstrate its potential as a physical attack.

Philipp Benz, Chaoning Zhang, Tooba Imtiaz et al.

The essence of deep learning is to exploit data to train a deep neural network (DNN) model. This work explores the reverse process of generating data from a model, attempting to reveal the relationship between the data and the model. We repeat the process of Data to Model (DtM) and Data from Model (DfM) in sequence and explore the loss of feature mapping information by measuring the accuracy drop on the original validation dataset. We perform this experiment for both a non-robust and robust origin model. Our results show that the accuracy drop is limited even after multiple sequences of DtM and DfM, especially for robust models. The success of this cycling transformation can be attributed to the shared feature mapping existing in data and model. Using the same data, we observe that different DtM processes result in models having different features, especially for different network architecture families, even though they achieve comparable performance.

Chaoning Zhang, Philipp Benz, Tooba Imtiaz et al.

A wide variety of works have explored the reason for the existence of adversarial examples, but there is no consensus on the explanation. We propose to treat the DNN logits as a vector for feature representation, and exploit them to analyze the mutual influence of two independent inputs based on the Pearson correlation coefficient (PCC). We utilize this vector representation to understand adversarial examples by disentangling the clean images and adversarial perturbations, and analyze their influence on each other. Our results suggest a new perspective towards the relationship between images and universal perturbations: Universal perturbations contain dominant features, and images behave like noise to them. This feature perspective leads to a new method for generating targeted universal adversarial perturbations using random source images. We are the first to achieve the challenging task of a targeted universal attack without utilizing original training data. Our approach using a proxy dataset achieves comparable performance to the state-of-the-art baselines which utilize the original training dataset.

Ho-Deok Jang, Sanghyun Woo, Philipp Benz et al.

We present a simple yet effective prediction module for a one-stage detector. The main process is conducted in a coarse-to-fine manner. First, the module roughly adjusts the default boxes to well capture the extent of target objects in an image. Second, given the adjusted boxes, the module aligns the receptive field of the convolution filters accordingly, not requiring any embedding layers. Both steps build a propose-and-attend mechanism, mimicking two-stage detectors in a highly efficient manner. To verify its effectiveness, we apply the proposed module to a basic one-stage detector SSD. Our final model achieves an accuracy comparable to that of state-of-the-art detectors while using a fraction of their model parameters and computational overheads. Moreover, we found that the proposed module has two strong applications. 1) The module can be successfully integrated into a lightweight backbone, further pushing the efficiency of the one-stage detector. 2) The module also allows train-from-scratch without relying on any sophisticated base networks as previous methods do.