Adversarial Robustness Against The Union Of Multiple Perturbation Models - How Is It Used In AI?
A lot of effort has been put into creating (empirically and certifiably) robust classifiers since deep learning systems are vulnerable to adversarial assaults. While most research has focused on protecting against a single attack type, several new studies have investigated adversarial robustness against the union of multiple perturbation models.
However, it may be challenging to adjust these approaches, and they can easily lead to uneven degrees of resilience to specific perturbation models, leading to a suboptimal worst-case loss across the union.
A logical modification of the traditional PGD-based technique is provided to combine several perturbation models into a single assault, assuming the worst-case scenario across all steepest descent directions. The best-case performance across the union is minimized.
Recent developments in adversarial training-based defenses have not yet rendered deep neural networks immune to adversarial assaults of a kind other than the perturbation type against which they have been taught to be resilient. A two-stage pipeline is used to increase resistance to a wide variety of disturbances.
COPYRIGHT_SZ: Published on https://stationzilla.com/adversarial-robustness-against-the-union-of-multiple-perturbation-models/ by Suleman Shah on 2022-11-07T11:23:19.273Z
When defending against a combined L1, L2, and assault, Protector surpasses previous adversarial training-based defenses by more than 5%. There is an inherent conflict between attacks on the top-level perturbation classifier and those on the second-level predictors: while strong attacks on the second-level predictors make it easier for the perturbation classifier to predict the adversarial perturbation type, fooling the perturbation classifier requires planting weaker (or less representative) attacks on the second-level predictors.
In this way, the model's overall resilience may be considerably improved by using even a flawed perturbation classifier. When compared to previous methods, Protector improves performance against the combined L1, L2, and L attacks by over 5%. To maximize adversarial accuracy against a certain class of attacks, such as norm-bounded perturbations, traditional adversarial training systems are narrow in scope.
Defenses against the union of several perturbations have been the focus of recent adversarial training expansions, although this advantage comes at the cost of a large (up to 10-fold) increase in training complexity over a single attack.
First, ResNet-50 and ResNet-101 are given a benchmark on ImageNet. Then, the adversarial accuracy of ResNet-18 on CIFAR-10 is improved by Shaped Noise Augmented Processing when it is put up against the union of (l, l2, and l1) perturbations.
Robustness against adversarial attacks and frequent perturbations are two separate concepts.
Training data is enhanced with perturbations borrowed from other models using the method of Ensemble Adversarial Training.
The degree to which a model's performance varies while incorporating fresh data vs training data is referred to as model resilience.
Adversarial robustness against the union of multiple perturbation models has the benefit of immediately convergent upon a trade-off between several perturbation models. By taking this approach, we are able to train standard architectures that are resilient against l, l2, and l1 attacks simultaneously, outperforming prior approaches on the MNIST and CIFAR10 datasets and improving upon adversarial accuracy of 40.6% on the latter by achieving 47.0% against the union of (l, l2, l1) perturbations with radius = (0.03, 0.5, 12).