Paper presents certified defenses against adversarial patch attacks.
problem Certified defenses against adversarial patch attacks are needed.
method Proposes the first certified defense and faster training methods.
result Demonstrates robustness transfer across different patch shapes.
This paper proposes a framework for certifying neural network defenses against data poisoning attacks.
problem Vulnerability of neural networks to data poisoning attacks.
method Random selection based defenses that average predictions on sub-datasets sampled from the training set.
result The certified radius of bagging derived by the framework is tighter than previous work.
Improved defense against data poisoning attacks by aggregating smaller subsets.
problem Mitigating the impact of poisoned data on model robustness.
method Finite Aggregation method that combines duplicates of smaller disjoint subsets for training.
result Consistent improvement in certified robustness bounds, up to 4.77% on GTSRB.
Deep Partition Aggregation defends against poisoning attacks with provable certificates.
problem Adversarial poisoning attacks corrupt classifier test-time behavior.
method Deep Partition Aggregation (DPA) is an ensemble method using hash partitions and base models trained on these partitions.
result DPA can certify >= 50% of test images against over 500 poison image insertions on MNIST, and nine insertions on CIFAR-10.
Defense against small image patches using occlusions.
problem Vulnerability of deep learning to small adversarial patches.
method Partially occlude image around each patch location.
result Certified security against patch attacks of a certain size.
SingleProp speeds up robust neural network training with minimal certification.
problem Efficiently defending neural networks against adversarial attacks with certified guarantees.
method SingleProp regularizer that requires only one forward pass per training iteration.
result Comparable certified accuracy to state-of-the-art defenses, but significantly faster training.
Automates perturbation analysis for neural networks, enabling certified robustness on complex architectures.
problem Limited applicability of existing perturbation analysis methods to complex neural network architectures.
method Developed an automatic framework to generalize LiRPA algorithms to any neural network structure, enabling loss fusion and state-of-the-art certified defense results.
result Demonstrated LiRPA based certified defense on Tiny ImageNet and Downscaled ImageNet.
Adversarial examples that fool machine learning models, particularly deep neural networks, have been a topic of intense research interest, with attacks and defenses being developed in a tight back-and-forth. Most past defenses are best effort and have been shown to be vulnerable to sophisticated attacks. Recently a set…
New method defends against patch attacks with high-certainty guarantees.
problem Patch attacks on images, especially physical adversarial attacks.
method Randomized smoothing, exploiting patch constraints.
result Meaningfully large robustness certificates against patch attacks.
How can we make machine learning provably robust against adversarial examples in a scalable way? Since certified defense methods, which ensure ε-robust, consume huge resources, they can only achieve small degree of robustness in practice. Lipschitz margin training (LMT) is a scalable certified defense, but it can als…
New attack tricks certifiably robust models into mislabeling images.
problem Defeating certified defenses against adversarial examples.
method Spoofed robustness certificates and large perturbations.
result Certifiably robust models can be fooled by large perturbations.
This paper provides a comprehensive benchmark and taxonomy for certifiably robust DNN defenses.
problem Certifiably robust defenses against adversarial attacks for deep neural networks.
method Taxonomy and benchmark of certifiably robust approaches.
result First comprehensive benchmark of certifiably robust approaches on different datasets.
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 advers…
Convolutional Neural Networks (CNNs) are deployed in more and more classification systems, but adversarial samples can be maliciously crafted to trick them, and are becoming a real threat. There have been various proposals to improve CNNs' adversarial robustness but these all suffer performance penalties or other limit…
New method certifies images against transformations like rotations and translations.
problem Certifying robustness of images against transformations like rotations and translations.
method Randomized smoothing with three different kinds of defenses.
result Individual certificates can be obtained via statistical error bounds or efficient online inverse computation.
This study improves scalability of randomized smoothing for certifying classifier robustness.
problem Certifying machine learning classifiers against adversarial attacks is challenging and scalable solutions are needed.
method The study reviews and explores randomized smoothing and its derivatives, focusing on scalability.
result The study provides theoretical guarantees and discusses scalability challenges of randomized smoothing.
Despite their impressive performance on diverse tasks, neural networks fail catastrophically in the presence of adversarial inputs---imperceptibly but adversarially perturbed versions of natural inputs. We have witnessed an arms race between defenders who attempt to train robust networks and attackers who try to constr…
The existence of adversarial data examples has drawn significant attention in the deep-learning community; such data are seemingly minimally perturbed relative to the original data, but lead to very different outputs from a deep-learning algorithm. Although a significant body of work on developing defensive models has …
New method certifies joint adversarial robustness of model ensembles.
problem Ensuring robustness of model ensembles against adversarial attacks.
method Proposes a novel technique to certify joint robustness, building on prior work on single-model robustness certification.
result Demonstrates the effectiveness of certifying joint robustness of ensembles, improving understanding of ensemble defenses.
BagCert efficiently certifies robustness against adversarial patches on image classifiers.
problem Adversarial patches pose a threat to autonomous systems' perception component.
method BagCert combines model architecture and certification procedure for efficient inference.
result BagCert certifies 10,000 examples in 43 seconds on a single GPU, achieving 86% clean and 60% certified accuracy against 5x5 patches.
This paper tackles robustness of ensemble stumps and trees under general ℓ_p norm perturbations.
problem The vulnerability of ensemble stumps and trees to small input perturbations under the ℓ_∞ norm.
method Developed dynamic programming algorithms for robustness verification and certified defense under general ℓ_p norm perturbations.
result First certified defense method for ensemble stumps and trees under ℓ_p norm perturbations.
New method certifies regression robustness without data distribution assumptions.
problem Certifying robustness of regression models against poisoning attacks.
method Reduces certified regression to certified classification using median decision function.
result Proposes six new provably-robust regression models.
We show how to turn any classifier that classifies well under Gaussian noise into a new classifier that is certifiably robust to adversarial perturbations under the ℓ2 norm. This "randomized smoothing" technique has been proposed recently in the literature, but existing guarantees are loose. We prove a tight robu…
Paper optimizes statistical estimation for randomized smoothing to reduce adversarial robustness certification time.
problem Efficiently estimating robustness of points against adversarial attacks.
method Developed estimation procedures using confidence sequences and randomized Clopper-Pearson intervals.
result Achieved optimal sample complexities and stronger certificates with reduced computational burden.
Several recent papers have discussed utilizing Lipschitz constants to limit the susceptibility of neural networks to adversarial examples. We analyze recently proposed methods for computing the Lipschitz constant. We show that the Lipschitz constant may indeed enable adversarially robust neural networks. However, the m…
Deep learning interpretation is essential to explain the reasoning behind model predictions. Understanding the robustness of interpretation methods is important especially in sensitive domains such as medical applications since interpretation results are often used in downstream tasks. Although gradient-based saliency …
Certified training improves robustness against adversarial attacks.
problem Certified training's gap with empirical robustness limits its practical utility.
method Combining adversarial attacks with network over-approximations.
result Certified training can prevent catastrophic overfitting and bridge the gap to multi-step baselines.
This work improves deep reinforcement learning robustness to adversarial state uncertainty.
problem Robustness of deep reinforcement learning to adversarial state uncertainty.
method Certified adversarial robustness techniques are applied to deep reinforcement learning algorithms to compute guaranteed lower bounds on state-action values.
result The approach increases robustness to noise and adversaries in pedestrian collision avoidance and classic control tasks.
Adversarial training is one of the most popular ways to learn robust models but is usually attack-dependent and time costly. In this paper, we propose the MACER algorithm, which learns robust models without using adversarial training but performs better than all existing provable l2-defenses. Recent work shows that ran…
Thanks to recent advances in deep neural networks (DNNs), face recognition systems have become highly accurate in classifying a large number of face images. However, recent studies have found that DNNs could be vulnerable to adversarial examples, raising concerns about the robustness of such systems. Adversarial exampl…
Improves safety region certification for smoothed classifiers without changing smoothing scheme.
problem Certified safety regions for smoothed classifiers are often small compared to optimal.
method Generalizes certified radius calculation as nested optimization problem, uses 0th-1st order information, and designs efficient estimators.
result Certified safety regions are significantly larger than current methods, achieving significant improvements on various metrics.
New method enhances neural network robustness against adversarial attacks.
problem Enhancing neural network robustness against adversarial attacks.
method Variational framework with per-sample noise level selector.
result Enhanced empirical robustness and certified robustness.
Neural networks are vulnerable to adversarial examples and researchers have proposed many heuristic attack and defense mechanisms. We address this problem through the principled lens of distributionally robust optimization, which guarantees performance under adversarial input perturbations. By considering a Lagrangian …
This paper assesses Gaussian and Exponential mechanisms for certifying adversarial robustness.
problem Certifying adversarial robustness using randomized smoothing mechanisms.
method Proposes a generic framework to assess the appropriateness of randomized smoothing mechanisms.
result Gaussian mechanism is an appropriate option for certifying both ℓ2-norm and ℓ∞-norm robustness. Aggregation defenses improve deep learning models' robustness against data poisoning attacks.
problem Data poisoning attacks manipulate deep learning models with malicious training samples.
method Deep Partition Aggregation, efficiency improvements, data-to-complexity ratio, poisoning overfitting phenomenon.
result Aggregation defenses boost poisoning robustness through the poisoning overfitting phenomenon.
Paper tackles fairness issues in GNNs by proposing ELEGANT for certification.
problem Fairness issues in GNN predictions due to graph data perturbations.
method Proposes ELEGANT framework for certifying fairness of any GNN without assumptions or re-training.
result The fairness of any GNN backbone is impossible to be corrupted under certain perturbation budgets.
Surveying strategies for making machine learning models robust against adversarial attacks.
problem Ensuring machine learning models are robust and reliable in real-world applications.
method Taxonomy of adversarial attacks and defenses, Robust Optimization problem formulation, and survey of methods.
result Surveyed recent results in adversarial example generation, defense mechanisms, and formal robustness certificates.
Robustness is an increasingly important property of machine learning models as they become more and more prevalent. We propose a defense against adversarial examples based on a k-nearest neighbor (kNN) on the intermediate activation of neural networks. Our scheme surpasses state-of-the-art defenses on MNIST and CIFAR-1…
Enhances robustness for time series classification using self-ensemble method.
problem Limited adversarial robustness in time series classification.
method Proposes a self-ensemble method to improve Randomized Smoothing's robustness certification.
result Demonstrates superior robustness compared to baseline approaches.
New neural network design resists small ℓ∞-norm adversarial perturbations.
problem Vulnerability of neural networks to small ℓ∞-norm adversarial perturbations. method Designing ℓ∞-dist neurons and constructing ℓ∞-dist nets, proving their 1-Lipschitz property and expressive power. result Certified robustness of ℓ∞-dist nets with state-of-the-art performance on various datasets. Paper defends deep learning classifiers against channel-aware adversarial attacks.
problem Deep learning classifiers are vulnerable to adversarial attacks.
method Channel-aware adversarial attacks are presented and defended against.
result Certified defense based on randomized smoothing makes classifiers robust.
Recently, techniques have been developed to provably guarantee the robustness of a classifier to adversarial perturbations of bounded L_1 and L_2 magnitudes by using randomized smoothing: the robust classification is a consensus of base classifications on randomly noised samples where the noise is additive. In this pap…
Paper proposes a robust metric learning algorithm.
problem Robustness of metric learning against adversarial perturbations is insufficient.
method Proposes a novel Mahalanobis distance metric learning algorithm.
result Certifiable robustness improvement over Euclidean distance.
Paper provides efficient robustness certificates for neural networks.
problem Ensuring neural networks are robust against adversarial attacks.
method Two-step approach: 1) Efficient convex optimization for robustness certificates with bounded Hessian eigenvalues, 2) Curvature-based regularization during training.
result Significantly higher certified robust accuracy achieved compared to existing methods.
The Lethal Dose Conjecture limits how much poisoned data can be tolerated.
problem Limiting the impact of poisoned data in machine learning models.
method Theoretical analysis and provable defenses (DPA, FA) based on majority voting.
result DPA and FA are asymptotically optimal defenses against data poisoning.
Adversarial examples are typically constructed by perturbing an existing data point within a small matrix norm, and current defense methods are focused on guarding against this type of attack. In this paper, we propose unrestricted adversarial examples, a new threat model where the attackers are not restricted to small…
Paper defends models from backdoor attacks using randomized smoothing.
problem Vulnerability of deep neural networks to backdoor attacks.
method Unified framework via randomized smoothing for both evasion and backdoor attacks; RAB robust training process.
result First provable robustness bound for backdoor attacks.
Training neural networks with verifiable robustness guarantees is challenging. Several existing approaches utilize linear relaxation based neural network output bounds under perturbation, but they can slow down training by a factor of hundreds depending on the underlying network architectures. Meanwhile, interval bound…