The pervasiveness of neural networks (NNs) in critical computer vision and image processing applications makes them very attractive for adversarial manipulation. A large body of existing research thoroughly investigates two broad categories of attacks targeting the integrity of NN models. The first category of attacks, commonly called Adversarial Examples, perturbs the model's inference by carefully adding noise into input examples. In the second category of attacks, adversaries try to manipulate the model during the training process by implanting Trojan backdoors. Researchers show that such attacks pose severe threats to the growing applications of NNs and propose several defenses against each attack type individually. However, such one-sided defense approaches leave potentially unknown risks in real-world scenarios when an adversary can unify different attacks to create new and more lethal ones bypassing existing defenses.
In this work, we show how to jointly exploit adversarial perturbation and model poisoning vulnerabilities to practically launch a new stealthy attack, dubbed AdvTrojan. AdvTrojan is stealthy because it can be activated only when: 1) a carefully crafted adversarial perturbation is injected into the input examples during inference, and 2) a Trojan backdoor is implanted during the training process of the model. We leverage adversarial noise in the input space to move Trojan-infected examples across the model decision boundary, making it difficult to detect. The stealthiness behavior of AdvTrojan fools the users into accidentally trusting the infected model as a robust classifier against adversarial examples. AdvTrojan can be implemented by only poisoning the training data similar to conventional Trojan backdoor attacks. Our thorough analysis and extensive experiments on several benchmark datasets show that AdvTrojan can bypass existing defenses with a success rate close to 100% in most of our experimental scenarios and can be extended to attack federated learning as well as high-resolution images.
more »
« less
An Invisible Black-Box Backdoor Attack Through Frequency Domain
Backdoor attacks have been shown to be a serious threat against deep learning systems such as biometric authentication and autonomous driving. An effective backdoor attack could enforce the model misbehave under certain predefined conditions, i.e., triggers, but behave normally otherwise. The triggers of existing attacks are mainly injected in the pixel space, which tend to be visually identifiable at both training and inference stages and detectable by existing defenses. In this paper, we propose a simple but effective and invisible black-box backdoor attack FTROJAN through trojaning the frequency domain. The key intuition is that triggering perturbations in the frequency domain correspond to small pixel-wise perturbations dispersed across the entire image, breaking the underlying assumptions of existing defenses and making the poisoning images visually indistinguishable from clean ones. Extensive experimental evaluations show that FTROJAN is highly effective and the poisoning images retain high perceptual quality. Moreover, we show that FTROJAN can robustly elude or significantly degenerate the performance of existing defenses.
more »
« less
- NSF-PAR ID:
- 10380827
- Date Published:
- Journal Name:
- ECCV2022
- Page Range / eLocation ID:
- 396–413
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
null (Ed.)In the last couple of years, several adversarial attack methods based on different threat models have been proposed for the image classification problem. Most existing defenses consider additive threat models in which sample perturbations have bounded L_p norms. These defenses, however, can be vulnerable against adversarial attacks under non-additive threat models. An example of an attack method based on a non-additive threat model is the Wasserstein adversarial attack proposed by Wong et al. (2019), where the distance between an image and its adversarial example is determined by the Wasserstein metric ("earth-mover distance") between their normalized pixel intensities. Until now, there has been no certifiable defense against this type of attack. In this work, we propose the first defense with certified robustness against Wasserstein Adversarial attacks using randomized smoothing. We develop this certificate by considering the space of possible flows between images, and representing this space such that Wasserstein distance between images is upper-bounded by L_1 distance in this flow-space. We can then apply existing randomized smoothing certificates for the L_1 metric. In MNIST and CIFAR-10 datasets, we find that our proposed defense is also practically effective, demonstrating significantly improved accuracy under Wasserstein adversarial attack compared to unprotected models.more » « less
-
Federated Learning (FL) allows individual clients to train a global model by aggregating local model updates each round. This results in collaborative model training while main-taining the privacy of clients' sensitive data. However, malicious clients can join the training process and train with poisoned data or send artificial model updates in targeted poisoning attacks. Many defenses to targeted poisoning attacks rely on anomaly-detection based metrics which remove participants that deviate from the majority. Similarly, aggregation-based defenses aim to reduce the impact of outliers, while L2-norm clipping tries to scale down the impact of malicious models. However, oftentimes these defenses misidentify benign clients as malicious or only work under specific attack conditions. In our paper, we examine the effectiveness of two anomaly -detection metrics on three different aggregation methods, in addition to the presence of L2-norm clipping and weight selection, across two different types of attacks. We also combine different defenses in order to examine their interaction and examine each defense when no attack is present. We found minimum aggregation to be the most effective defense against label-flipping attacks, whereas both minimum aggregation and geometric median worked well against distributed backdoor attacks. Using random weight selection significantly deteriorated defenses against both attacks, whereas the use of clipping made little difference. Finally, the main task accuracy was directly correlated with the BA in the label-flipping attack and generally was close to the MA in benign scenarios. However, in the DBA the MA and BA are inversely correlated and the MA fluctuates greatly.more » « less
-
Existing backdoor defense methods are only effective for limited trigger types. To defend different trigger types at once, we start from the class-irrelevant nature of the poisoning process and propose a novel weakly supervised backdoor defense framework WeDef. Recent advances in weak supervision make it possible to train a reasonably accurate text classifier using only a small number of user-provided, class-indicative seed words. Such seed words shall be considered independent of the triggers. Therefore, a weakly supervised text classifier trained by only the poisoned documents without their labels will likely have no backdoor. Inspired by this observation, in WeDef, we define the reliability of samples based on whether the predictions of the weak classifier agree with their labels in the poisoned training set. We further improve the results through a two-phase sanitization: (1) iteratively refine the weak classifier based on the reliable samples and (2) train a binary poison classifier by distinguishing the most unreliable samples from the most reliable samples. Finally, we train the sanitized model on the samples that the poison classifier predicts as benign. Extensive experiments show that WeDef is effective against popular trigger-based attacks (e.g., words, sentences, and paraphrases), outperforming existing defense methods.more » « less
-
The prevalence of data scraping from social media as a means to obtain datasets has led to growing concerns regarding unauthorized use of data. Data poisoning attacks have been proposed as a bulwark against scraping, as they make data "unlearnable'' by adding small, imperceptible perturbations. Unfortunately, existing methods require knowledge of both the target architecture and the complete dataset so that a surrogate network can be trained, the parameters of which are used to generate the attack. In this work, we introduce autoregressive (AR) poisoning, a method that can generate poisoned data without access to the broader dataset. The proposed AR perturbations are generic, can be applied across different datasets, and can poison different architectures. Compared to existing unlearnable methods, our AR poisons are more resistant against common defenses such as adversarial training and strong data augmentations. Our analysis further provides insight into what makes an effective data poison.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
