Jian Wang

Xiaofei Xie, Tianlin Li, Jian Wang et al.

Deep learning has recently been widely applied to many applications across different domains, e.g., image classification and audio recognition. However, the quality of Deep Neural Networks (DNNs) still raises concerns in the practical operational environment, which calls for systematic testing, especially in safety-critical scenarios. Inspired by software testing, a number of structural coverage criteria are designed and proposed to measure the test adequacy of DNNs. However, due to the blackbox nature of DNN, the existing structural coverage criteria are difficult to interpret, making it hard to understand the underlying principles of these criteria. The relationship between the structural coverage and the decision logic of DNNs is unknown. Moreover, recent studies have further revealed the non-existence of correlation between the structural coverage and DNN defect detection, which further posts concerns on what a suitable DNN testing criterion should be. In this paper, we propose the interpretable coverage criteria through constructing the decision structure of a DNN. Mirroring the control flow graph of the traditional program, we first extract a decision graph from a DNN based on its interpretation, where a path of the decision graph represents a decision logic of the DNN. Based on the control flow and data flow of the decision graph, we propose two variants of path coverage to measure the adequacy of the test cases in exercising the decision logic. The higher the path coverage, the more diverse decision logic the DNN is expected to be explored. Our large-scale evaluation results demonstrate that: the path in the decision graph is effective in characterizing the decision of the DNN, and the proposed coverage criteria are also sensitive with errors including natural errors and adversarial examples, and strongly correlated with the output impartiality.

Wenkai Tan, Justus Renkhoff, Alvaro Velasquez et al.

Deep Learning (DL) and Deep Neural Networks (DNNs) are widely used in various domains. However, adversarial attacks can easily mislead a neural network and lead to wrong decisions. Defense mechanisms are highly preferred in safety-critical applications. In this paper, firstly, we use the gradient class activation map (GradCAM) to analyze the behavior deviation of the VGG-16 network when its inputs are mixed with adversarial perturbation or Gaussian noise. In particular, our method can locate vulnerable layers that are sensitive to adversarial perturbation and Gaussian noise. We also show that the behavior deviation of vulnerable layers can be used to detect adversarial examples. Secondly, we propose a novel NoiseCAM algorithm that integrates information from globally and pixel-level weighted class activation maps. Our algorithm is susceptible to adversarial perturbations and will not respond to Gaussian random noise mixed in the inputs. Third, we compare detecting adversarial examples using both behavior deviation and NoiseCAM, and we show that NoiseCAM outperforms behavior deviation modeling in its overall performance. Our work could provide a useful tool to defend against certain adversarial attacks on deep neural networks.

Justus Renkhoff, Wenkai Tan, Alvaro Velasquez et al.

Deep Learning (DL) is being applied in various domains, especially in safety-critical applications such as autonomous driving. Consequently, it is of great significance to ensure the robustness of these methods and thus counteract uncertain behaviors caused by adversarial attacks. In this paper, we use gradient heatmaps to analyze the response characteristics of the VGG-16 model when the input images are mixed with adversarial noise and statistically similar Gaussian random noise. In particular, we compare the network response layer by layer to determine where errors occurred. Several interesting findings are derived. First, compared to Gaussian random noise, intentionally generated adversarial noise causes severe behavior deviation by distracting the area of concentration in the networks. Second, in many cases, adversarial examples only need to compromise a few intermediate blocks to mislead the final decision. Third, our experiments revealed that specific blocks are more vulnerable and easier to exploit by adversarial examples. Finally, we demonstrate that the layers $Block4\_conv1$ and $Block5\_cov1$ of the VGG-16 model are more susceptible to adversarial attacks. Our work could provide valuable insights into developing more reliable Deep Neural Network (DNN) models.

Chak Tou Leong, Yi Cheng, Jiashuo Wang et al.

Language model detoxification aims to minimize the risk of generating offensive or harmful content in pretrained language models (PLMs) for safer deployment. Existing methods can be roughly categorized as finetuning-based and decoding-based. However, the former is often resource-intensive, while the latter relies on additional components and potentially compromises the generation fluency. In this paper, we propose a more lightweight approach that enables the PLM itself to achieve "self-detoxification". Our method is built upon the observation that prepending a negative steering prompt can effectively induce PLMs to generate toxic content. At the same time, we are inspired by the recent research in the interpretability field, which formulates the evolving contextualized representations within the PLM as an information stream facilitated by the attention layers. Drawing on this idea, we devise a method to identify the toxification direction from the normal generation process to the one prompted with the negative prefix, and then steer the generation to the reversed direction by manipulating the information movement within the attention layers. Experimental results show that our approach, without any fine-tuning or extra components, can achieve comparable performance with state-of-the-art methods.

Qing Guo, Felix Juefei-Xu, Xiaofei Xie et al.

The state-of-the-art deep neural networks (DNNs) are vulnerable against adversarial examples with additive random-like noise perturbations. While such examples are hardly found in the physical world, the image blurring effect caused by object motion, on the other hand, commonly occurs in practice, making the study of which greatly important especially for the widely adopted real-time image processing tasks (e.g., object detection, tracking). In this paper, we initiate the first step to comprehensively investigate the potential hazards of the blur effect for DNN, caused by object motion. We propose a novel adversarial attack method that can generate visually natural motion-blurred adversarial examples, named motion-based adversarial blur attack (ABBA). To this end, we first formulate the kernel-prediction-based attack where an input image is convolved with kernels in a pixel-wise way, and the misclassification capability is achieved by tuning the kernel weights. To generate visually more natural and plausible examples, we further propose the saliency-regularized adversarial kernel prediction, where the salient region serves as a moving object, and the predicted kernel is regularized to achieve naturally visual effects. Besides, the attack is further enhanced by adaptively tuning the translations of object and background. A comprehensive evaluation on the NeurIPS'17 adversarial competition dataset demonstrates the effectiveness of ABBA by considering various kernel sizes, translations, and regions. The in-depth study further confirms that our method shows more effective penetrating capability to the state-of-the-art GAN-based deblurring mechanisms compared with other blurring methods. We release the code to https://github.com/tsingqguo/ABBA.

Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo et al.

Segment Anything Model (SAM) has emerged as a transformative approach in image segmentation, acclaimed for its robust zero-shot segmentation capabilities and flexible prompting system. Nonetheless, its performance is challenged by images with degraded quality. Addressing this limitation, we propose the Robust Segment Anything Model (RobustSAM), which enhances SAM's performance on low-quality images while preserving its promptability and zero-shot generalization. Our method leverages the pre-trained SAM model with only marginal parameter increments and computational requirements. The additional parameters of RobustSAM can be optimized within 30 hours on eight GPUs, demonstrating its feasibility and practicality for typical research laboratories. We also introduce the Robust-Seg dataset, a collection of 688K image-mask pairs with different degradations designed to train and evaluate our model optimally. Extensive experiments across various segmentation tasks and datasets confirm RobustSAM's superior performance, especially under zero-shot conditions, underscoring its potential for extensive real-world application. Additionally, our method has been shown to effectively improve the performance of SAM-based downstream tasks such as single image dehazing and deblurring.

Mingfu Xue, Shifeng Ni, Yinghao Wu et al.

Recent researches demonstrate that Deep Neural Networks (DNN) models are vulnerable to backdoor attacks. The backdoored DNN model will behave maliciously when images containing backdoor triggers arrive. To date, existing backdoor attacks are single-trigger and single-target attacks, and the triggers of most existing backdoor attacks are obvious thus are easy to be detected or noticed. In this paper, we propose a novel imperceptible and multi-channel backdoor attack against Deep Neural Networks by exploiting Discrete Cosine Transform (DCT) steganography. Based on the proposed backdoor attack method, we implement two variants of backdoor attacks, i.e., N-to-N backdoor attack and N-to-One backdoor attack. Specifically, for a colored image, we utilize DCT steganography to construct the trigger on different channels of the image. As a result, the trigger is stealthy and natural. Based on the proposed method, we implement multi-target and multi-trigger backdoor attacks. Experimental results demonstrate that the average attack success rate of the N-to-N backdoor attack is 93.95% on CIFAR-10 dataset and 91.55% on TinyImageNet dataset, respectively. The average attack success rate of N-to-One attack is 90.22% and 89.53% on CIFAR-10 and TinyImageNet datasets, respectively. Meanwhile, the proposed backdoor attack does not affect the classification accuracy of the DNN model. Moreover, the proposed attack is demonstrated to be robust to the state-of-the-art backdoor defense (Neural Cleanse).

Mingfu Xue, Xin Wang, Shichang Sun et al.

In recent years, many backdoor attacks based on training data poisoning have been proposed. However, in practice, those backdoor attacks are vulnerable to image compressions. When backdoor instances are compressed, the feature of specific backdoor trigger will be destroyed, which could result in the backdoor attack performance deteriorating. In this paper, we propose a compression-resistant backdoor attack based on feature consistency training. To the best of our knowledge, this is the first backdoor attack that is robust to image compressions. First, both backdoor images and their compressed versions are input into the deep neural network (DNN) for training. Then, the feature of each image is extracted by internal layers of the DNN. Next, the feature difference between backdoor images and their compressed versions are minimized. As a result, the DNN treats the feature of compressed images as the feature of backdoor images in feature space. After training, the backdoor attack against DNN is robust to image compression. Furthermore, we consider three different image compressions (i.e., JPEG, JPEG2000, WEBP) in feature consistency training, so that the backdoor attack is robust to multiple image compression algorithms. Experimental results demonstrate the effectiveness and robustness of the proposed backdoor attack. When the backdoor instances are compressed, the attack success rate of common backdoor attack is lower than 10%, while the attack success rate of our compression-resistant backdoor is greater than 97%. The compression-resistant attack is still robust even when the backdoor images are compressed with low compression quality. In addition, extensive experiments have demonstrated that, our compression-resistant backdoor attack has the generalization ability to resist image compression which is not used in the training process.

Mingfu Xue, Yinghao Wu, Yushu Zhang et al.

Training high performance Deep Neural Networks (DNNs) models require large-scale and high-quality datasets. The expensive cost of collecting and annotating large-scale datasets make the valuable datasets can be considered as the Intellectual Property (IP) of the dataset owner. To date, almost all the copyright protection schemes for deep learning focus on the copyright protection of models, while the copyright protection of the dataset is rarely studied. In this paper, we propose a novel method to actively protect the dataset from being used to train DNN models without authorization. Experimental results on on CIFAR-10 and TinyImageNet datasets demonstrate the effectiveness of the proposed method. Compared with the model trained on clean dataset, the proposed method can effectively make the test accuracy of the unauthorized model trained on protected dataset drop from 86.21% to 38.23% and from 74.00% to 16.20% on CIFAR-10 and TinyImageNet datasets, respectively.

Haoqi Wang, Mingfu Xue, Shichang Sun et al.

Deep neural networks (DNN) have achieved remarkable performance in various fields. However, training a DNN model from scratch requires a lot of computing resources and training data. It is difficult for most individual users to obtain such computing resources and training data. Model copyright infringement is an emerging problem in recent years. For instance, pre-trained models may be stolen or abuse by illegal users without the authorization of the model owner. Recently, many works on protecting the intellectual property of DNN models have been proposed. In these works, embedding watermarks into DNN based on backdoor is one of the widely used methods. However, when the DNN model is stolen, the backdoor-based watermark may face the risk of being detected and removed by an adversary. In this paper, we propose a scheme to detect and remove watermark in deep neural networks via generative adversarial networks (GAN). We demonstrate that the backdoor-based DNN watermarks are vulnerable to the proposed GAN-based watermark removal attack. The proposed attack method includes two phases. In the first phase, we use the GAN and few clean images to detect and reverse the watermark in the DNN model. In the second phase, we fine-tune the watermarked DNN based on the reversed backdoor images. Experimental evaluations on the MNIST and CIFAR10 datasets demonstrate that, the proposed method can effectively remove about 98% of the watermark in DNN models, as the watermark retention rate reduces from 100% to less than 2% after applying the proposed attack. In the meantime, the proposed attack hardly affects the model's performance. The test accuracy of the watermarked DNN on the MNIST and the CIFAR10 datasets drops by less than 1% and 3%, respectively.

Mingfu Xue, Zhiyu Wu, Jian Wang et al.

A well-trained DNN model can be regarded as an intellectual property (IP) of the model owner. To date, many DNN IP protection methods have been proposed, but most of them are watermarking based verification methods where model owners can only verify their ownership passively after the copyright of DNN models has been infringed. In this paper, we propose an effective framework to actively protect the DNN IP from infringement. Specifically, we encrypt the DNN model's parameters by perturbing them with well-crafted adversarial perturbations. With the encrypted parameters, the accuracy of the DNN model drops significantly, which can prevent malicious infringers from using the model. After the encryption, the positions of encrypted parameters and the values of the added adversarial perturbations form a secret key. Authorized user can use the secret key to decrypt the model. Compared with the watermarking methods which only passively verify the ownership after the infringement occurs, the proposed method can prevent infringement in advance. Moreover, compared with most of the existing active DNN IP protection methods, the proposed method does not require additional training process of the model, which introduces low computational overhead. Experimental results show that, after the encryption, the test accuracy of the model drops by 80.65%, 81.16%, and 87.91% on Fashion-MNIST, CIFAR-10, and GTSRB, respectively. Moreover, the proposed method only needs to encrypt an extremely low number of parameters, and the proportion of the encrypted parameters of all the model's parameters is as low as 0.000205%. The experimental results also indicate that, the proposed method is robust against model fine-tuning attack and model pruning attack. Moreover, for the adaptive attack where attackers know the detailed steps of the proposed method, the proposed method is also demonstrated to be robust.

Shichang Sun, Mingfu Xue, Jian Wang et al.

Recently, the research on protecting the intellectual properties (IP) of deep neural networks (DNN) has attracted serious concerns. A number of DNN copyright protection methods have been proposed. However, most of the existing watermarking methods focus on verifying the copyright of the model, which do not support the authentication and management of users' fingerprints, thus can not satisfy the requirements of commercial copyright protection. In addition, the query modification attack which was proposed recently can invalidate most of the existing backdoor-based watermarking methods. To address these challenges, in this paper, we propose a method to protect the intellectual properties of DNN models by using an additional class and steganographic images. Specifically, we use a set of watermark key samples to embed an additional class into the DNN, so that the watermarked DNN will classify the watermark key sample as the predefined additional class in the copyright verification stage. We adopt the least significant bit (LSB) image steganography to embed users' fingerprints into watermark key images. Each user will be assigned with a unique fingerprint image so that the user's identity can be authenticated later. Experimental results demonstrate that, the proposed method can protect the copyright of DNN models effectively. On Fashion-MNIST and CIFAR-10 datasets, the proposed method can obtain 100% watermark accuracy and 100% fingerprint authentication success rate. In addition, the proposed method is demonstrated to be robust to the model fine-tuning attack, model pruning attack, and the query modification attack. Compared with three existing watermarking methods (the logo-based, noise-based, and adversarial frontier stitching watermarking methods), the proposed method has better performance on watermark accuracy and robustness against the query modification attack.

Mingfu Xue, Can He, Shichang Sun et al.

Deep neural networks (DNN) have been widely deployed in various applications. However, many researches indicated that DNN is vulnerable to backdoor attacks. The attacker can create a hidden backdoor in target DNN model, and trigger the malicious behaviors by submitting specific backdoor instance. However, almost all the existing backdoor works focused on the digital domain, while few studies investigate the backdoor attacks in real physical world. Restricted to a variety of physical constraints, the performance of backdoor attacks in the real physical world will be severely degraded. In this paper, we propose a robust physical backdoor attack method, PTB (physical transformations for backdoors), to implement the backdoor attacks against deep learning models in the real physical world. Specifically, in the training phase, we perform a series of physical transformations on these injected backdoor instances at each round of model training, so as to simulate various transformations that a backdoor may experience in real world, thus improves its physical robustness. Experimental results on the state-of-the-art face recognition model show that, compared with the backdoor methods that without PTB, the proposed attack method can significantly improve the performance of backdoor attacks in real physical world. Under various complex physical conditions, by injecting only a very small ratio (0.5%) of backdoor instances, the attack success rate of physical backdoor attacks with the PTB method on VGGFace is 82%, while the attack success rate of backdoor attacks without the proposed PTB method is lower than 11%. Meanwhile, the normal performance of the target DNN model has not been affected.

Lebin Yu, Jian Wang, Xudong Zhang

Reinforcement learning has achieved remarkable performance in a wide range of tasks these days. Nevertheless, some unsolved problems limit its applications in real-world control. One of them is model misspecification, a situation where an agent is trained and deployed in environments with different transition dynamics. We propose an novel framework that utilize history trajectory and Partial Observable Markov Decision Process Modeling to deal with this dilemma. Additionally, we put forward an efficient adversarial attack method to assist robust training. Our experiments in four gym domains validate the effectiveness of our framework.

Mingfu Xue, Shichang Sun, Can He et al.

The training of Deep Neural Networks (DNN) is costly, thus DNN can be considered as the intellectual properties (IP) of model owners. To date, most of the existing protection works focus on verifying the ownership after the DNN model is stolen, which cannot resist piracy in advance. To this end, we propose an active DNN IP protection method based on adversarial examples against DNN piracy, named ActiveGuard. ActiveGuard aims to achieve authorization control and users' fingerprints management through adversarial examples, and can provide ownership verification. Specifically, ActiveGuard exploits the elaborate adversarial examples as users' fingerprints to distinguish authorized users from unauthorized users. Legitimate users can enter fingerprints into DNN for identity authentication and authorized usage, while unauthorized users will obtain poor model performance due to an additional control layer. In addition, ActiveGuard enables the model owner to embed a watermark into the weights of DNN. When the DNN is illegally pirated, the model owner can extract the embedded watermark and perform ownership verification. Experimental results show that, for authorized users, the test accuracy of LeNet-5 and Wide Residual Network (WRN) models are 99.15% and 91.46%, respectively, while for unauthorized users, the test accuracy of the two DNNs are only 8.92% (LeNet-5) and 10% (WRN), respectively. Besides, each authorized user can pass the fingerprint authentication with a high success rate (up to 100%). For ownership verification, the embedded watermark can be successfully extracted, while the normal performance of the DNN model will not be affected. Further, ActiveGuard is demonstrated to be robust against fingerprint forgery attack, model fine-tuning attack and pruning attack.