Abstract:The neural network memorization problem is to study the expressive power of neural networks to interpolate a finite dataset. Although memorization is widely believed to have a close relationship with the strong generalizability of deep learning when using over-parameterized models, to the best of our knowledge, there exists no theoretical study on the generalizability of memorization neural networks. In this paper, we give the first theoretical analysis of this topic. Since using i.i.d. training data is a necessary condition for a learning algorithm to be generalizable, memorization and its generalization theory for i.i.d. datasets are developed under mild conditions on the data distribution. First, algorithms are given to construct memorization networks for an i.i.d. dataset, which have the smallest number of parameters and even a constant number of parameters. Second, we show that, in order for the memorization networks to be generalizable, the width of the network must be at least equal to the dimension of the data, which implies that the existing memorization networks with an optimal number of parameters are not generalizable. Third, a lower bound for the sample complexity of general memorization algorithms and the exact sample complexity for memorization algorithms with constant number of parameters are given. It is also shown that there exist data distributions such that, to be generalizable for them, the memorization network must have an exponential number of parameters in the data dimension. Finally, an efficient and generalizable memorization algorithm is given when the number of training samples is greater than the efficient memorization sample complexity of the data distribution.
Abstract:In recent years, the study of adversarial robustness in object detection systems, particularly those based on deep neural networks (DNNs), has become a pivotal area of research. Traditional physical attacks targeting object detectors, such as adversarial patches and texture manipulations, directly manipulate the surface of the object. While these methods are effective, their overt manipulation of objects may draw attention in real-world applications. To address this, this paper introduces a more subtle approach: an inconspicuous adversarial trigger that operates outside the bounding boxes, rendering the object undetectable to the model. We further enhance this approach by proposing the Feature Guidance (FG) technique and the Universal Auto-PGD (UAPGD) optimization strategy for crafting high-quality triggers. The effectiveness of our method is validated through extensive empirical testing, demonstrating its high performance in both digital and physical environments. The code and video will be available at: https://github.com/linToTao/Out-of-bbox-attack.
Abstract:The recent development of Sora leads to a new era in text-to-video (T2V) generation. Along with this comes the rising concern about its security risks. The generated videos may contain illegal or unethical content, and there is a lack of comprehensive quantitative understanding of their safety, posing a challenge to their reliability and practical deployment. Previous evaluations primarily focus on the quality of video generation. While some evaluations of text-to-image models have considered safety, they cover fewer aspects and do not address the unique temporal risk inherent in video generation. To bridge this research gap, we introduce T2VSafetyBench, a new benchmark designed for conducting safety-critical assessments of text-to-video models. We define 12 critical aspects of video generation safety and construct a malicious prompt dataset using LLMs and jailbreaking prompt attacks. Based on our evaluation results, we draw several important findings, including: 1) no single model excels in all aspects, with different models showing various strengths; 2) the correlation between GPT-4 assessments and manual reviews is generally high; 3) there is a trade-off between the usability and safety of text-to-video generative models. This indicates that as the field of video generation rapidly advances, safety risks are set to surge, highlighting the urgency of prioritizing video safety. We hope that T2VSafetyBench can provide insights for better understanding the safety of video generation in the era of generative AI.
Abstract:The generalization bound is a crucial theoretical tool for assessing the generalizability of learning methods and there exist vast literatures on generalizability of normal learning, adversarial learning, and data poisoning. Unlike other data poison attacks, the backdoor attack has the special property that the poisoned triggers are contained in both the training set and the test set and the purpose of the attack is two-fold. To our knowledge, the generalization bound for the backdoor attack has not been established. In this paper, we fill this gap by deriving algorithm-independent generalization bounds in the clean-label backdoor attack scenario. Precisely, based on the goals of backdoor attack, we give upper bounds for the clean sample population errors and the poison population errors in terms of the empirical error on the poisoned training dataset. Furthermore, based on the theoretical result, a new clean-label backdoor attack is proposed that computes the poisoning trigger by combining adversarial noise and indiscriminate poison. We show its effectiveness in a variety of settings.
Abstract:Privacy preserving has become increasingly critical with the emergence of social media. Unlearnable examples have been proposed to avoid leaking personal information on the Internet by degrading generalization abilities of deep learning models. However, our study reveals that unlearnable examples are easily detectable. We provide theoretical results on linear separability of certain unlearnable poisoned dataset and simple network based detection methods that can identify all existing unlearnable examples, as demonstrated by extensive experiments. Detectability of unlearnable examples with simple networks motivates us to design a novel defense method. We propose using stronger data augmentations coupled with adversarial noises generated by simple networks, to degrade the detectability and thus provide effective defense against unlearnable examples with a lower cost. Adversarial training with large budgets is a widely-used defense method on unlearnable examples. We establish quantitative criteria between the poison and adversarial budgets which determine the existence of robust unlearnable examples or the failure of the adversarial defense.
Abstract:Invariance to spatial transformations such as translations and rotations is a desirable property and a basic design principle for classification neural networks. However, the commonly used convolutional neural networks (CNNs) are actually very sensitive to even small translations. There exist vast works to achieve exact or approximate transformation invariance by designing transformation-invariant models or assessing the transformations. These works usually make changes to the standard CNNs and harm the performance on standard datasets. In this paper, rather than modifying the classifier, we propose a pre-classifier restorer to recover translated (or even rotated) inputs to the original ones which will be fed into any classifier for the same dataset. The restorer is based on a theoretical result which gives a sufficient and necessary condition for an affine operator to be translational equivariant on a tensor space.
Abstract:Adversarial deep learning is to train robust DNNs against adversarial attacks, which is one of the major research focuses of deep learning. Game theory has been used to answer some of the basic questions about adversarial deep learning such as the existence of a classifier with optimal robustness and the existence of optimal adversarial samples for a given class of classifiers. In most previous work, adversarial deep learning was formulated as a simultaneous game and the strategy spaces are assumed to be certain probability distributions in order for the Nash equilibrium to exist. But, this assumption is not applicable to the practical situation. In this paper, we give answers to these basic questions for the practical case where the classifiers are DNNs with a given structure, by formulating the adversarial deep learning as sequential games. The existence of Stackelberg equilibria for these games are proved. Furthermore, it is shown that the equilibrium DNN has the largest adversarial accuracy among all DNNs with the same structure, when Carlini-Wagner's margin loss is used. Trade-off between robustness and accuracy in adversarial deep learning is also studied from game theoretical aspect.
Abstract:In this paper, a new parameter perturbation attack on DNNs, called adversarial parameter attack, is proposed, in which small perturbations to the parameters of the DNN are made such that the accuracy of the attacked DNN does not decrease much, but its robustness becomes much lower. The adversarial parameter attack is stronger than previous parameter perturbation attacks in that the attack is more difficult to be recognized by users and the attacked DNN gives a wrong label for any modified sample input with high probability. The existence of adversarial parameters is proved. For a DNN $F_{\Theta}$ with the parameter set $\Theta$ satisfying certain conditions, it is shown that if the depth of the DNN is sufficiently large, then there exists an adversarial parameter set $\Theta_a$ for $\Theta$ such that the accuracy of $F_{\Theta_a}$ is equal to that of $F_{\Theta}$, but the robustness measure of $F_{\Theta_a}$ is smaller than any given bound. An effective training algorithm is given to compute adversarial parameters and numerical experiments are used to demonstrate that the algorithms are effective to produce high quality adversarial parameters.
Abstract:In this paper, the bias classifier is introduced, that is, the bias part of a DNN with Relu as the activation function is used as a classifier. The work is motivated by the fact that the bias part is a piecewise constant function with zero gradient and hence cannot be directly attacked by gradient-based methods to generate adversaries such as FGSM. The existence of the bias classifier is proved an effective training method for the bias classifier is proposed. It is proved that by adding a proper random first-degree part to the bias classifier, an information-theoretically safe classifier against the original-model gradient-based attack is obtained in the sense that the attack generates a totally random direction for generating adversaries. This seems to be the first time that the concept of information-theoretically safe classifier is proposed. Several attack methods for the bias classifier are proposed and numerical experiments are used to show that the bias classifier is more robust than DNNs against these attacks in most cases.
Abstract:In this paper, we present a robust classification-autoencoder (CAE) which has strong ability to recognize outliers and defend adversaries. The basic idea is to change the autoencoder from an unsupervised learning method into a classifier. The CAE is a modified autoencoder, where the encoder is used to compress samples with different labels into disjoint compression spaces and the decoder is used to recover a sample with a given label from the corresponding compression space. The encoder is used as a classifier and the decoder is used to decide whether the classification given by the encoder is correct by comparing the input sample with the output. Since adversary samples are seeming inevitable for the current DNN framework, we introduce the list classification based on CAE to defend adversaries, which outputs several labels and the corresponding samples recovered by the CAE. The CAE is evaluated using the MNIST dataset in great detail. It is shown that the CAE network can recognize almost all outliers and the list classification contains the correct label for almost all adversaries.