Backdoor Attacks

Vision Transformers (ViTs), which made a splash in the field of computer vision (CV), have shaken the dominance of convolutional neural networks (CNNs). However, in the process of industrializing ViTs, backdoor attacks have brought severe challenges to security. The success of ViTs benefits from the self-attention mechanism. However, compared with CNNs, we find that this mechanism of capturing global information within patches makes ViTs more sensitive to patch-wise triggers. Under such observations, we delicately design a novel backdoor attack framework for ViTs, dubbed BadViT, which utilizes a universal patch-wise trigger to catch the model's attention from patches beneficial for classification to those with triggers, thereby manipulating the mechanism on which ViTs survive to confuse itself. Furthermore, we propose invisible variants of BadViT to increase the stealth of the attack by limiting the strength of the trigger perturbation. Through a large number of experiments, it is proved that BadViT is an efficient backdoor attack method against ViTs, which is less dependent on the number of poisons, with satisfactory convergence, and is transferable for downstream tasks. Furthermore, the risks inside of ViTs to backdoor attacks are also explored from the perspective of existing advanced defense schemes.

Diffusion models have achieved great success in a range of tasks, such as image synthesis and molecule design. As such successes hinge on large-scale training data collected from diverse sources, the trustworthiness of these collected data is hard to control or audit. In this work, we aim to explore the vulnerabilities of diffusion models under potential training data manipulations and try to answer: How hard is it to perform Trojan attacks on well-trained diffusion models? What are the adversarial targets that such Trojan attacks can achieve? To answer these questions, we propose an effective Trojan attack against diffusion models, TrojDiff, which optimizes the Trojan diffusion and generative processes during training. In particular, we design novel transitions during the Trojan diffusion process to diffuse adversarial targets into a biased Gaussian distribution and propose a new parameterization of the Trojan generative process that leads to an effective training objective for the attack. In addition, we consider three types of adversarial targets: the Trojaned diffusion models will always output instances belonging to a certain class from the in-domain distribution (In-D2D attack), out-of-domain distribution (Out-D2D-attack), and one specific instance (D2I attack). We evaluate TrojDiff on CIFAR-10 and CelebA datasets against both DDPM and DDIM diffusion models. We show that TrojDiff always achieves high attack performance under different adversarial targets using different types of triggers, while the performance in benign environments is preserved. The code is available at https://github.com/chenweixin107/TrojDiff.

Recent advancements in deploying deep neural networks (DNNs) on resource-constrained devices have generated interest in input-adaptive dynamic neural networks (DyNNs). DyNNs offer more efficient inferences and enable the deployment of DNNs on devices with limited resources, such as mobile devices. However, we have discovered a new vulnerability in DyNNs that could potentially compromise their efficiency. Specifically, we investigate whether adversaries can manipulate DyNNs' computational costs to create a false sense of efficiency. To address this question, we propose EfficFrog, an adversarial attack that injects universal efficiency backdoors in DyNNs. To inject a backdoor trigger into DyNNs, EfficFrog poisons only a minimal percentage of the DyNNs' training data. During the inference phase, EfficFrog can slow down the backdoored DyNNs and abuse the computational resources of systems running DyNNs by adding the trigger to any input. To evaluate EfficFrog, we tested it on three DNN backbone architectures (based on VGG16, MobileNet, and ResNet56) using two popular datasets (CIFAR-10 and Tiny ImageNet). Our results demonstrate that EfficFrog reduces the efficiency of DyNNs on triggered input samples while keeping the efficiency of clean samples almost the same.

Many defenses against adversarial attacks (e.g. robust classifiers, randomization, or image purification) use countermeasures put to work only after the attack has been crafted. We adopt a different perspective to introduce A^5 (Adversarial Augmentation Against Adversarial Attacks), a novel framework including the first certified preemptive defense against adversarial attacks. The main idea is to craft a defensive perturbation to guarantee that any attack (up to a given magnitude) towards the input in hand will fail. To this aim, we leverage existing automatic perturbation analysis tools for neural networks. We study the conditions to apply A^5 effectively, analyze the importance of the robustness of the to-be-defended classifier, and inspect the appearance of the robustified images. We show effective on-the-fly defensive augmentation with a robustifier network that ignores the ground truth label, and demonstrate the benefits of robustifier and classifier co-training. In our tests, A^5 consistently beats state of the art certified defenses on MNIST, CIFAR10, FashionMNIST and Tinyimagenet. We also show how to apply A^5 to create certifiably robust physical objects. The released code at https://github.com/NVlabs/A5 allows experimenting on a wide range of scenarios beyond the man-in-the-middle attack tested here, including the case of physical attacks.

Backdoor inversion, the process of finding a backdoor trigger inserted into a machine learning model, has become the pillar of many backdoor detection and defense methods. Previous works on backdoor inversion often recover the backdoor through an optimization process to flip a support set of clean images into the target class. However, it is rarely studied and understood how large this support set should be to recover a successful backdoor. In this work, we show that one can reliably recover the backdoor trigger with as few as a single image. Specifically, we propose the SmoothInv method, which first constructs a robust smoothed version of the backdoored classifier and then performs guided image synthesis towards the target class to reveal the backdoor pattern. SmoothInv requires neither an explicit modeling of the backdoor via a mask variable, nor any complex regularization schemes, which has become the standard practice in backdoor inversion methods. We perform both quantitaive and qualitative study on backdoored classifiers from previous published backdoor attacks. We demonstrate that compared to existing methods, SmoothInv is able to recover successful backdoors from single images, while maintaining high fidelity to the original backdoor. We also show how we identify the target backdoored class from the backdoored classifier. Last, we propose and analyze two countermeasures to our approach and show that SmoothInv remains robust in the face of an adaptive attacker. Our code is available at https://github.com/locuslab/smoothinv.

A hard challenge in developing practical face recognition (FR) attacks is due to the black-box nature of the target FR model, i.e., inaccessible gradient and parameter information to attackers. While recent research took an important step towards attacking black-box FR models through leveraging transferability, their performance is still limited, especially against online commercial FR systems that can be pessimistic (e.g., a less than 50% ASR--attack success rate on average). Motivated by this, we present Sibling-Attack, a new FR attack technique for the first time explores a novel multi-task perspective (i.e., leveraging extra information from multi-correlated tasks to boost attacking transferability). Intuitively, Sibling-Attack selects a set of tasks correlated with FR and picks the Attribute Recognition (AR) task as the task used in Sibling-Attack based on theoretical and quantitative analysis. Sibling-Attack then develops an optimization framework that fuses adversarial gradient information through (1) constraining the cross-task features to be under the same space, (2) a joint-task meta optimization framework that enhances the gradient compatibility among tasks, and (3) a cross-task gradient stabilization method which mitigates the oscillation effect during attacking. Extensive experiments demonstrate that Sibling-Attack outperforms state-of-the-art FR attack techniques by a non-trivial margin, boosting ASR by 12.61% and 55.77% on average on state-of-the-art pre-trained FR models and two well-known, widely used commercial FR systems.

Model inversion attacks are a type of privacy attack that reconstructs private data used to train a machine learning model, solely by accessing the model. Recently, white-box model inversion attacks leveraging Generative Adversarial Networks (GANs) to distill knowledge from public datasets have been receiving great attention because of their excellent attack performance. On the other hand, current black-box model inversion attacks that utilize GANs suffer from issues such as being unable to guarantee the completion of the attack process within a predetermined number of query accesses or achieve the same level of performance as white-box attacks. To overcome these limitations, we propose a reinforcement learning-based black-box model inversion attack. We formulate the latent space search as a Markov Decision Process (MDP) problem and solve it with reinforcement learning. Our method utilizes the confidence scores of the generated images to provide rewards to an agent. Finally, the private data can be reconstructed using the latent vectors found by the agent trained in the MDP. The experiment results on various datasets and models demonstrate that our attack successfully recovers the private information of the target model by achieving state-of-the-art attack performance. We emphasize the importance of studies on privacy-preserving machine learning by proposing a more advanced black-box model inversion attack.

Model inversion (MI) attacks aim to infer and reconstruct private training data by abusing access to a model. MI attacks have raised concerns about the leaking of sensitive information (e.g. private face images used in training a face recognition system). Recently, several algorithms for MI have been proposed to improve the attack performance. In this work, we revisit MI, study two fundamental issues pertaining to all state-of-the-art (SOTA) MI algorithms, and propose solutions to these issues which lead to a significant boost in attack performance for all SOTA MI. In particular, our contributions are two-fold: 1) We analyze the optimization objective of SOTA MI algorithms, argue that the objective is sub-optimal for achieving MI, and propose an improved optimization objective that boosts attack performance significantly. 2) We analyze "MI overfitting", show that it would prevent reconstructed images from learning semantics of training data, and propose a novel "model augmentation" idea to overcome this issue. Our proposed solutions are simple and improve all SOTA MI attack accuracy significantly. E.g., in the standard CelebA benchmark, our solutions improve accuracy by 11.8% and achieve for the first time over 90% attack accuracy. Our findings demonstrate that there is a clear risk of leaking sensitive information from deep learning models. We urge serious consideration to be given to the privacy implications. Our code, demo, and models are available at https://ngoc-nguyen-0.github.io/re-thinking_model_inversion_attacks/

Deep neural networks (DNNs) are known to be vulnerable to both backdoor attacks as well as adversarial attacks. In the literature, these two types of attacks are commonly treated as distinct problems and solved separately, since they belong to training-time and inference-time attacks respectively. However, in this paper we find an intriguing connection between them: for a model planted with backdoors, we observe that its adversarial examples have similar behaviors as its triggered samples, i.e., both activate the same subset of DNN neurons. It indicates that planting a backdoor into a model will significantly affect the model's adversarial examples. Based on this observations, a novel Progressive Backdoor Erasing (PBE) algorithm is proposed to progressively purify the infected model by leveraging untargeted adversarial attacks. Different from previous backdoor defense methods, one significant advantage of our approach is that it can erase backdoor even when the additional clean dataset is unavailable. We empirically show that, against 5 state-of-the-art backdoor attacks, our AFT can effectively erase the backdoor triggers without obvious performance degradation on clean samples and significantly outperforms existing defense methods.

We develop a novel method to remove injected backdoors in deep learning models. It works by cloning the benign behaviors of a trojaned model to a new model of the same structure. It trains the clone model from scratch on a very small subset of samples and aims to minimize a cloning loss that denotes the differences between the activations of important neurons across the two models. The set of important neurons varies for each input, depending on their magnitude of activations and their impact on the classification result. We theoretically show our method can better recover benign functions of the backdoor model. Meanwhile, we prove our method can be more effective in removing backdoors compared with fine-tuning. Our experiments show that our technique can effectively remove nine different types of backdoors with minor benign accuracy degradation, outperforming the state-of-the-art backdoor removal techniques that are based on fine-tuning, knowledge distillation, and neuron pruning.

Diffusion models are state-of-the-art deep learning empowered generative models that are trained based on the principle of learning forward and reverse diffusion processes via progressive noise-addition and denoising. To gain a better understanding of the limitations and potential risks, this paper presents the first study on the robustness of diffusion models against backdoor attacks. Specifically, we propose BadDiffusion, a novel attack framework that engineers compromised diffusion processes during model training for backdoor implantation. At the inference stage, the backdoored diffusion model will behave just like an untampered generator for regular data inputs, while falsely generating some targeted outcome designed by the bad actor upon receiving the implanted trigger signal. Such a critical risk can be dreadful for downstream tasks and applications built upon the problematic model. Our extensive experiments on various backdoor attack settings show that BadDiffusion can consistently lead to compromised diffusion models with high utility and target specificity. Even worse, BadDiffusion can be made cost-effective by simply finetuning a clean pre-trained diffusion model to implant backdoors. We also explore some possible countermeasures for risk mitigation. Our results call attention to potential risks and possible misuse of diffusion models.

We propose an approach for adversarial attacks on dense prediction models (such as object detectors and segmentation). It is well known that the attacks generated by a single surrogate model do not transfer to arbitrary (blackbox) victim models. Furthermore, targeted attacks are often more challenging than the untargeted attacks. In this paper, we show that a carefully designed ensemble can create effective attacks for a number of victim models. In particular, we show that normalization of the weights for individual models plays a critical role in the success of the attacks. We then demonstrate that by adjusting the weights of the ensemble according to the victim model can further improve the performance of the attacks. We performed a number of experiments for object detectors and segmentation to highlight the significance of the our proposed methods. Our proposed ensemble-based method outperforms existing blackbox attack methods for object detection and segmentation. Finally we show that our proposed method can also generate a single perturbation that can fool multiple blackbox detection and segmentation models simultaneously.

In this paper, we propose a novel transfer-based targeted attack method that optimizes the adversarial perturbations without any extra training efforts for auxiliary networks on training data. Our new attack method is proposed based on the observation that highly universal adversarial perturbations tend to be more transferable for targeted attacks. Therefore, we propose to make the perturbation to be agnostic to different local regions within one image, which we called as self-universality. Instead of optimizing the perturbations on different images, optimizing on different regions to achieve self-universality can get rid of using extra data. Specifically, we introduce a feature similarity loss that encourages the learned perturbations to be universal by maximizing the feature similarity between adversarial perturbed global images and randomly cropped local regions. With the feature similarity loss, our method makes the features from adversarial perturbations to be more dominant than that of benign images, hence improving targeted transferability. We name the proposed attack method as Self-Universality (SU) attack. Extensive experiments demonstrate that SU can achieve high success rates for transfer-based targeted attacks. On ImageNet-compatible dataset, SU yields an improvement of 12% compared with existing state-of-the-art methods. Code is available at https://github.com/zhipeng-wei/Self-Universality.

Attackers can deceive neural networks by adding human imperceptive perturbations to their input data; this reveals the vulnerability and weak robustness of current deep-learning networks. Many attack techniques have been proposed to evaluate the model's robustness. Gradient-based attacks suffer from severely overestimating the robustness. This paper identifies that the relative error in calculated gradients caused by floating-point errors, including floating-point underflow and rounding errors, is a fundamental reason why gradient-based attacks fail to accurately assess the model's robustness. Although it is hard to eliminate the relative error in the gradients, we can control its effect on the gradient-based attacks. Correspondingly, we propose an efficient loss function by minimizing the detrimental impact of the floating-point errors on the attacks. Experimental results show that it is more efficient and reliable than other loss functions when examined across a wide range of defence mechanisms.

Adversarial attacks can evaluate model robustness and have been of great concerns in recent years. Among various attacks, targeted attacks aim at misleading victim models to output adversary-desired predictions, which are more challenging and threatening than untargeted ones. Existing targeted attacks can be roughly divided into instancespecific and instance-agnostic attacks. Instance-specific attacks craft adversarial examples via iterative gradient updating on the specific instance. In contrast, instanceagnostic attacks learn a universal perturbation or a generative model on the global dataset to perform attacks. However they rely too much on the classification boundary of substitute models, ignoring the realistic distribution of target class, which may result in limited targeted attack performance. And there is no attempt to simultaneously combine the information of the specific instance and the global dataset. To deal with these limitations, we first conduct an analysis via a causal graph and propose to craft transferable targeted adversarial examples by injecting target patterns. Based on this analysis, we introduce a generative attack model composed of a cross-attention guided convolution module and a pattern injection module. Concretely, the former adopts a dynamic convolution kernel and a static convolution kernel for the specific instance and the global dataset, respectively, which can inherit the advantages of both instance-specific and instance-agnostic attacks. And the pattern injection module utilizes a pattern prototype to encode target patterns, which can guide the generation of targeted adversarial examples. Besides, we also provide rigorous theoretical analysis to guarantee the effectiveness of our method. Extensive experiments demonstrate that our method show superior performance than 10 existing adversarial attacks against 13 models.

Self-supervised learning in computer vision trains on unlabeled data, such as images or (image, text) pairs, to obtain an image encoder that learns high-quality embeddings for input data. Emerging backdoor attacks towards encoders expose crucial vulnerabilities of self-supervised learning, since downstream classifiers (even further trained on clean data) may inherit backdoor behaviors from encoders. Existing backdoor detection methods mainly focus on supervised learning settings and cannot handle pre-trained encoders especially when input labels are not available. In this paper, we propose DECREE, the first backdoor detection approach for pre-trained encoders, requiring neither classifier headers nor input labels. We evaluate DECREE on over 400 encoders trojaned under 3 paradigms. We show the effectiveness of our method on image encoders pre-trained on ImageNet and OpenAI's CLIP 400 million image-text pairs. Our method consistently has a high detection accuracy even if we have only limited or no access to the pre-training dataset.

Deep neural networks are proven to be vulnerable to backdoor attacks. Detecting the trigger samples during the inference stage, i.e., the test-time trigger sample detection, can prevent the backdoor from being triggered. However, existing detection methods often require the defenders to have high accessibility to victim models, extra clean data, or knowledge about the appearance of backdoor triggers, limiting their practicality. In this paper, we propose the test-time corruption robustness consistency evaluation (TeCo), a novel test-time trigger sample detection method that only needs the hard-label outputs of the victim models without any extra information. Our journey begins with the intriguing observation that the backdoor-infected models have similar performance across different image corruptions for the clean images, but perform discrepantly for the trigger samples. Based on this phenomenon, we design TeCo to evaluate test-time robustness consistency by calculating the deviation of severity that leads to predictions' transition across different corruptions. Extensive experiments demonstrate that compared with state-of-the-art defenses, which even require either certain information about the trigger types or accessibility of clean data, TeCo outperforms them on different backdoor attacks, datasets, and model architectures, enjoying a higher AUROC by 10% and 5 times of stability. The code is available at https://github.com/CGCL-codes/TeCo

Recently, self-supervised learning (SSL) was shown to be vulnerable to patch-based data poisoning backdoor attacks. It was shown that an adversary can poison a small part of the unlabeled data so that when a victim trains an SSL model on it, the final model will have a backdoor that the adversary can exploit. This work aims to defend self-supervised learning against such attacks. We use a three-step defense pipeline, where we first train a model on the poisoned data. In the second step, our proposed defense algorithm (PatchSearch) uses the trained model to search the training data for poisoned samples and removes them from the training set. In the third step, a final model is trained on the cleaned-up training set. Our results show that PatchSearch is an effective defense. As an example, it improves a model's accuracy on images containing the trigger from 38.2% to 63.7% which is very close to the clean model's accuracy, 64.6%. Moreover, we show that PatchSearch outperforms baselines and state-of-the-art defense approaches including those using additional clean, trusted data. Our code is available at https://github.com/UCDvision/PatchSearch

Backdoor attacks against neural networks have been intensively investigated, where the adversary compromises the integrity of the victim model, causing it to make wrong predictions for inference samples containing a specific trigger. To make the trigger more imperceptible and human-unnoticeable, a variety of stealthy backdoor attacks have been proposed, some works employ imperceptible perturbations as the backdoor triggers, which restrict the pixel differences of the triggered image and clean image. Some works use special image styles (e.g., reflection, Instagram filter) as the backdoor triggers. However, these attacks sacrifice the robustness, and can be easily defeated by common preprocessing-based defenses. This paper presents a novel color backdoor attack, which can exhibit robustness and stealthiness at the same time. The key insight of our attack is to apply a uniform color space shift for all pixels as the trigger. This global feature is robust to image transformation operations and the triggered samples maintain natural-looking. To find the optimal trigger, we first define naturalness restrictions through the metrics of PSNR, SSIM and LPIPS. Then we employ the Particle Swarm Optimization (PSO) algorithm to search for the optimal trigger that can achieve high attack effectiveness and robustness while satisfying the restrictions. Extensive experiments demonstrate the superiority of PSO and the robustness of color backdoor against different mainstream backdoor defenses.