Proposes a new adversarial model to avoid accuracy vs. adversarial accuracy tradeoff.
problem Inherent tradeoff between accuracy and adversarial accuracy in existing adversarial robustness definitions.
method Introduces Voronoi-epsilon adversary that balances perturbation constraints.
result Voronoi-epsilon adversary avoids accuracy vs. adversarial accuracy tradeoff even with large ε. Adversarial consistency depends on the uniqueness of adversarial Bayes classifiers.
problem Consistency of adversarial surrogate losses is not guaranteed.
method Connected consistency of adversarial surrogate losses to the uniqueness of adversarial Bayes classifiers.
result A convex surrogate loss is statistically consistent for adversarial learning if and only if the adversarial Bayes classifier is unique.
New method makes neural networks more resilient to location-optimized adversarial patches.
problem Neural networks' vulnerability to adversarial patches that are visible but still effective.
method Developed a practical approach to optimize patch locations and applied adversarial training.
result Significantly improved robustness against adversarial patches on CIFAR10 and GTSRB.
Adversarial training achieves optimal test error for shallow networks.
problem Achieving optimal adversarial test error for general data distributions.
method Applying new Rademacher complexity bounds and properties of optimal adversarial predictors.
result Adversarial training can achieve optimal adversarial test error for general data distributions.
SPAT improves adversarial robustness by preserving semantics in adversarial training.
problem Adversarial examples often have different semantics than original data, introducing unintended biases.
method Semantics-preserving adversarial training (SPAT) that encourages pixel perturbation shared among all classes.
result SPAT improves adversarial robustness and achieves state-of-the-art results in CIFAR-10 and CIFAR-100.
Adversarial training, in which a network is trained on both adversarial and clean examples, is one of the most trusted defense methods against adversarial attacks. However, there are three major practical difficulties in implementing and deploying this method - expensive in terms of extra memory and computation costs; …
By injecting adversarial examples into training data, adversarial training is promising for improving the robustness of deep learning models. However, most existing adversarial training approaches are based on a specific type of adversarial attack. It may not provide sufficiently representative samples from the adversa…
Paper bounds convergence rate of adversarial surrogate risk.
problem Vulnerability of binary classification models to adversarial attacks.
method Characterizes conditions for adversarial consistency and provides surrogate risk bounds.
result Surrogate risk bounds quantify the rate of convergence of adversarial classification risk.
Study on robustness in linear regression models, focusing on adversarial perturbations.
problem Understanding and improving robustness in linear regression models to adversarial perturbations.
method Developed a two-stage adversarial learning framework that incorporates model structure information.
result Proved the consistency and developed the Bahadur representation of the adversarially robust estimator.
Adversarial training based on the minimax formulation is necessary for obtaining adversarial robustness of trained models. However, it is conservative or even pessimistic so that it sometimes hurts the natural generalization. In this paper, we raise a fundamental question---do we have to trade off natural generalizatio…
Ideally, what confuses neural network should be confusing to humans. However, recent experiments have shown that small, imperceptible perturbations can change the network prediction. To address this gap in perception, we propose a novel approach for learning robust classifier. Our main idea is: adversarial examples for…
A Bayesian framework models adversarial uncertainty for robust machine learning.
problem Vulnerability of machine learning models to adversarial attacks.
method Formal Bayesian framework that models adversarial uncertainty through a stochastic channel, articulating probabilistic assumptions.
result Explicitly modeling adversarial uncertainty leads to improved robustification strategies.
Modern machine learning algorithms perform poorly on adversarially manipulated data. Adversarial risk quantifies the error of classifiers in adversarial settings; adversarial classifiers minimize adversarial risk. In this paper, we analyze adversarial risk and adversarial classifiers from an optimal transport perspecti…
Adversarial training improves audio classifier robustness.
problem Improving robustness of audio classifiers against adversarial attacks.
method Adversarial training of deep neural networks on audio data.
result Adversarially trained models can resist higher perturbations and fooling rates.
Neural networks are vulnerable to adversarial attacks -- small visually imperceptible crafted noise which when added to the input drastically changes the output. The most effective method of defending against these adversarial attacks is to use the methodology of adversarial training. We analyze the adversarially train…
Paper tackles robustness in adversarial noise with a meta-optimizer.
problem Sensitivity to adversarial noise hinders machine learning deployment.
method Meta-optimizer learns to robustly optimize models using adversarial examples.
result Meta-optimizer transfers adversarial knowledge to new models without generating new examples.
New adversarial examples with structured distortion sets improve robustness and perceptibility.
problem Improving adversarial robustness and perceptibility of images.
method Exploring and constraining adversarial search with trace-norms and other norms.
result Structured adversarial perturbations allow larger distortions and control over generation.
Robust RL with learned optimal adversary improves agent performance under adversarial state observations.
problem Ensuring reinforcement learning agents' robustness against adversarial perturbations of state observations.
method Proposed a framework of alternating training with learned adversaries (ATLA) to find optimal adversarial policies and enhance agent robustness.
result ATLA achieves state-of-the-art performance under strong adversaries in continuous control environments.
In recent years, deep neural networks have demonstrated outstanding performance in many machine learning tasks. However, researchers have discovered that these state-of-the-art models are vulnerable to adversarial examples: legitimate examples added by small perturbations which are unnoticeable to human eyes. Adversari…
The paper studies the asymptotic behavior of adversarial training under ℓ∞-perturbation.
problem Theoretical guarantees for sparsity-recovery in adversarial training.
method Investigation of the asymptotic distribution of the adversarial training estimator in generalized linear models.
result The asymptotic distribution of the adversarial training estimator under ℓ∞-perturbation could have a positive probability mass at 0 when the true parameter is 0. Adversarial training is one of the strongest defenses against adversarial attacks, but it requires adversarial examples to be generated for every mini-batch during optimization. The expense of producing these examples during training often precludes adversarial training from use on complex image datasets. In this study…
This paper explores tradeoffs between standard and adversarial risks in distributionally adversarial training.
problem Understanding the impact of adversarial training on standard risk and adversarial risk.
method Study of distributionally adversarial training with different learning settings and models.
result Derives Pareto-optimal tradeoff curves between standard and adversarial risks.
Paper proposes a new black-box adversarial attack using normalizing flows.
problem Adversarial vulnerability of deep neural networks.
method Proposes a novel black-box adversarial attack using normalizing flows.
result Demonstrates competitive performance against well-known black-box adversarial attack methods.
This paper improves adversarial robustness of deep learning models.
problem Vulnerability of machine learning models to adversarial perturbations.
method Analyzes adversarial training for linear regression and neural networks, incorporating L1 penalty.
result Incorporating L1 penalty leads to consistent adversarially robust estimation in high-dimensional settings.
Framework to mitigate adversarial attacks by allowing classifiers to abstain.
problem Vulnerability of classifiers to adversarial examples.
method Adversarial training with a rejection option.
result ATRO framework improves classifier reliability against adversarial attacks.
Today's state-of-the-art image classifiers fail to correctly classify carefully manipulated adversarial images. In this work, we develop a new, localized adversarial attack that generates adversarial examples by imperceptibly altering the backgrounds of normal images. We first use this attack to highlight the unnecessa…
A new method reduces adversarial training time without overfitting.
problem Catastrophic overfitting in single-step adversarial training.
method FGSMPR: FGSM with PGD Regularization.
result Reduces the gap to multi-step adversarial training.
Adversarial samples are perturbed inputs crafted to mislead the machine learning systems. A training mechanism, called adversarial training, which presents adversarial samples along with clean samples has been introduced to learn robust models. In order to scale adversarial training for large datasets, these perturbati…
I describe an optimal control view of adversarial machine learning, where the dynamical system is the machine learner, the input are adversarial actions, and the control costs are defined by the adversary's goals to do harm and be hard to detect. This view encompasses many types of adversarial machine learning, includi…
This paper investigates how network width and depth affect adversarially robust DNNs.
problem Understanding architectural configurations for adversarially robust DNNs.
method Comprehensive investigation on the impact of network width and depth on adversarial robustness.
result Optimal architectural configuration for adversarial robustness exists and can improve robustness.
Study non-asymptotic bounds for robust estimators under misspecified models.
problem Evaluate performance of robust estimators under adversarial conditions.
method Propose a general approach to adversarial risk analysis, including investigations on generalization and approximation errors.
result Establish non-asymptotic upper bounds for adversarial excess risk under Lipschitz loss functions.
3D adversarial logos can fool object detectors in real-world settings.
problem Creating robust adversarial attacks in 3D rendering views.
method Constructing 3D adversarial logos via texture mapping and differentiable rendering.
result 3D adversarial logos are more versatile and robust than traditional adversarial patches.
Adversarial robustness has become a central goal in deep learning, both in the theory and the practice. However, successful methods to improve the adversarial robustness (such as adversarial training) greatly hurt generalization performance on the unperturbed data. This could have a major impact on how the adversarial …
Adversarial training is a training scheme designed to counter adversarial attacks by augmenting the training dataset with adversarial examples. Surprisingly, several studies have observed that loss gradients from adversarially trained DNNs are visually more interpretable than those from standard DNNs. Although this phe…
Paper tackles catastrophic overfitting in single-step adversarial training.
problem Catastrophic overfitting leads to sudden drop in robust accuracy.
method Proposes a method to prevent overfitting by using all adversarial examples.
result Demonstrates prevention of catastrophic overfitting and improves robustness.
The paper examines how neural network topology affects adversarial robustness.
problem Understanding how neural network topology influences adversarial robustness.
method Investigated the graph of input traversing all layers of a neural network, comparing clean and adversarial inputs.
result Under-optimized edges in neural network graphs are a source of adversarial vulnerability and can be used to detect adversarial inputs.
Study analyzes adversarial training dynamics without data distribution assumptions.
problem Understanding training dynamics of adversarial training without data distribution assumptions.
method Mean field theory approach to analyze adversarial training in random deep neural networks.
result Upper bounds of adversarial loss derived empirically and theoretically.
Paper explores fast adversarial training to improve robustness with less computation.
problem Efficiently defending against adversarial examples.
method Integrates simple self-attacks for faster training, focusing on overfitting recovery.
result Shows superior robust accuracy with reduced training time compared to strong adversarial training.
Proposes a probabilistic method for generating semantically-aware adversarial examples.
problem Generating adversarial examples that are difficult for humans to detect while preserving semantics.
method Embeds subjective understanding of semantics as a distribution into adversarial example generation.
result Achieves higher success rates in circumventing adversarial defense mechanisms.
This study connects Jacobian regularization to adversarial robustness and improves generalization.
problem Adversarial attacks make deep neural networks vulnerable.
method Developed a connection between Jacobian regularization and adversarial training, and established robust generalization gaps.
result Jacobian norms are related to both standard and robust generalization.
Bayesian adversaries can outsmart traditional adversarial attacks.
problem Bayesian adversaries can manipulate machine learning models through small perturbations.
method Developed a continuous-time particle system (Abram) to approximate the gradient flow of the Bayesian adversarial robustness problem.
result Abram approximates the minimizer of the Bayesian adversarial robustness problem under certain assumptions.
Paper generates diverse, readable adversarial texts from scratch.
problem Text classification models are easily fooled by adversarial examples.
method Trained a conditional variational autoencoder (VAE) with adversarial loss and utilized GANs to generate consistent adversarial texts.
result Successfully generates adversarial texts with higher success rate and acceptable quality.
We study two important concepts in adversarial deep learning---adversarial training and generative adversarial network (GAN). Adversarial training is the technique used to improve the robustness of discriminator by combining adversarial attacker and discriminator in the training phase. GAN is commonly used for image ge…
Adversarial training, in which a network is trained on adversarial examples, is one of the few defenses against adversarial attacks that withstands strong attacks. Unfortunately, the high cost of generating strong adversarial examples makes standard adversarial training impractical on large-scale problems like ImageNet…
Analyzes adversarial training's impact on loss landscape, proposing PAS to improve model performance.
problem Challenges in optimizing models under adversarial training due to loss landscape properties.
method Analytical studies of adversarial loss functions, numerical analyses, PAS strategy.
result Adversarial training impairs optimization, but PAS strategy improves model performance.
BPN defends against adversarial attacks by generating beneficial perturbations.
problem Adversarial attacks cause deep neural networks to misclassify clean inputs.
method BPN generates beneficial perturbations during training to neutralize future adversarial attacks.
result BPN is robust to adversarial examples and more efficient than classical adversarial training.
RTFE provides adversarial robustness to multiple models.
problem Adversarial examples can transfer to other models, compromising robustness.
method Proposes RTFE, a deep learning-based pre-processing mechanism.
result RTFE provides adversarial robustness to multiple independently trained classifiers.
New measure assesses deep neural networks' robustness to adversarial attacks.
problem Deep learning's fragility to adversarial attacks limits its adoption in mission-critical applications.
method Introduces residual error as a new performance measure for assessing adversarial robustness.
result Demonstrates effectiveness of residual error in assessing robustness of deep neural networks.