Enforcing Linearity in DNN succours Robustness and Adversarial Image Generation

Sarkar, Anindya; Gupta, Nikhil; Iyengar, Raghu Sesha

doi:10.48550/arxiv.1910.08108

Cited by 2 publications

(4 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Accuracy, mean±std% [18] 85.31 ± 0.41 53.42 ± 1.07 He et al [16] 82.84 ± 0.22 46.11 ± 0.43 Balaji et al [4] 91.34 48.53 Sarkar et al [31] 87.65 54.77 Smooth (our) 81.07 ± 0.01 58.56 ± 0.14 a results by He et al [16] racy on clean samples, taking into account the quality of the base classifier and amount of Monte Carlo samples.…”

Section: Methodsmentioning

confidence: 98%

“…Xie et al [39] proposed a feature denoising mechanism inspired by self-attention [33] and non-local blocks [35]. Sarkar et al [31] introduced loss terms which enforce linear structure and Lipschitz continuity of the network.…”

Section: Black Box Attacksmentioning

confidence: 99%

“…Deep neural networks (DNNs) are showing spectacular performance in a variety of computer vision tasks, but at the same time are susceptible to adversarial examples -small perturbations that alter the prediction [14,32]. Since the initial discovery of this phenomenon in 2013, growingly stronger defences [14,14,16,21,26,30,31,39,41] and counter-attacks [3,6,9,14,22,26,28,29] were proposed in the literature. Adversarial attacks have also been shown in tasks beyond image classification where they were first discovered, object detection [34], natural language process- * Equal contribution.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Smoothed Inference for Adversarially-Trained Models

Nemcovsky,

Zheltonozhskii,

Baskin

et al. 2019

Preprint

View full text Add to dashboard Cite

Deep neural networks are known to be vulnerable to inputs with maliciously constructed adversarial perturbations aimed at forcing misclassification. We study randomized smoothing as a way to both improve performance on unperturbed data as well as increase robustness to adversarial attacks. Moreover, we extend the method proposed by He et al. [16] by adding low-rank multivariate noise, which we then use as a base model for smoothing. The proposed method achieves 58.5% top-1 accuracy on CIFAR-10 under PGD attack and outperforms previous works by 4%. In addition, we consider a family of attacks, which were previously used for training purposes in the certified robustness scheme. We demonstrate that the proposed attacks are more effective than PGD against both smoothed and non-smoothed models. Since our method is based on sampling, it lends itself well for trading-off between the model inference complexity and its performance. A reference implementation of the proposed techniques is provided.

show abstract

Section: Methodsmentioning

confidence: 98%

Section: Black Box Attacksmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Smoothed Inference for Adversarially-Trained Models

Nemcovsky,

Zheltonozhskii,

Baskin

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…Adversarial attacks were discovered in the seminal paper by Szegedy et al [47], and have since been extensively explored in the image domain [17,29,52,44,24,58,20,55,5,2,29,43,36,8,26], natural language processing [14,6,22], and reinforcement learning [16], to name just a few. In this paper, we focus on universal attacks for geometric data, hence this section covers relevant prior work addressing the two aspects.…”

Section: Related Workmentioning

confidence: 99%

Universal Spectral Adversarial Attacks for Deformable Shapes

Rampini¹,

Pestarini²,

Cosmo³

et al. 2021

Preprint

View full text Add to dashboard Cite

Machine learning models are known to be vulnerable to adversarial attacks, namely perturbations of the data that lead to wrong predictions despite being imperceptible. However, the existence of "universal" attacks (i.e., unique perturbations that transfer across different data points) has only been demonstrated for images to date. Part of the reason lies in the lack of a common domain, for geometric data such as graphs, meshes, and point clouds, where a universal perturbation can be defined. In this paper, we offer a change in perspective and demonstrate the existence of universal attacks for geometric data (shapes). We introduce a computational procedure that operates entirely in the spectral domain, where the attacks take the form of small perturbations to short eigenvalue sequences; the resulting geometry is then synthesized via shape-from-spectrum recovery. Our attacks are universal, in that they transfer across different shapes, different representations (meshes and point clouds), and generalize to previously unseen data.

show abstract

Enforcing Linearity in DNN succours Robustness and Adversarial Image Generation

Cited by 2 publications

References 0 publications

Smoothed Inference for Adversarially-Trained Models

Smoothed Inference for Adversarially-Trained Models

Universal Spectral Adversarial Attacks for Deformable Shapes

Contact Info

Product

Resources

About