Search

Deep neural networks (DNNs) have gained prominence in various applications, but remain vulnerable to adversarial attacks that manipulate data to mislead a DNN. This paper aims to challenge the efficacy and transferability of two contemporary defense mechanisms against adversarial attacks: (a) robust training and (b) adversarial training. The former suggests that training a DNN on a data set consisting solely of robust features should produce a model resistant to adversarial attacks. The latter creates an adversarially trained model that learns to minimise an expected training loss over a distribution of bounded adversarial perturbations. We reveal a significant lack in the transferability of these defense mechanisms and provide insight into the potential dangers posed by L∞-norm attacks previously underestimated by the research community. Such conclusions are based on extensive experiments involving (1) different model architectures, (2) the use of canonical correlation analysis, (3) visual and quantitative analysis of the neural network’s latent representations, (4) an analysis of networks’ decision boundaries and (5) the use of equivalence of L2 and L∞ perturbation norm theories.

Deep neural networks are exposed to the risk of adversarial attacks via the fast gradient sign method (FGSM), projected gradient descent (PGD) attacks, and other attack algorithms. Adversarial training is one of the methods used to defend against the threat of adversarial attacks. It is a training schema that utilizes an alternative objective function to provide model generalization for both adversarial data and clean data. In this systematic review, we focus particularly on adversarial training as a method of improving the defensive capacities and robustness of machine learning models. Specifically, we focus on adversarial sample accessibility through adversarial sample generation methods. The purpose of this systematic review is to survey state-of-the-art adversarial training and robust optimization methods to identify the research gaps within this field of applications. The literature search was conducted using Engineering Village (Engineering Village is an engineering literature search tool, which provides access to 14 engineering literature and patent databases), where we collected 238 related papers. The papers were filtered according to defined inclusion and exclusion criteria, and information was extracted from these papers according to a defined strategy. A total of 78 papers published between 2016 and 2021 were selected. Data were extracted and categorized using a defined strategy, and bar plots and comparison tables were used to show the data distribution. The findings of this review indicate that there are limitations to adversarial training methods and robust optimization. The most common problems are related to data generalization and overfitting.

Modern trainable image recognition models are vulnerable to different types of perturbations; hence, the development of resilient intelligent algorithms for safety-critical applications remains a relevant concern to reduce the impact of perturbation on model performance. This paper proposes a model and training method for a resilient image classifier capable of efficiently functioning despite various faults, adversarial attacks, and concept drifts. The proposed model has a multi-section structure with a hierarchy of optimized class prototypes and hyperspherical class boundaries, which provides adaptive computation, perturbation absorption, and graceful degradation. The proposed training method entails the application of a complex loss function assembled from its constituent parts in a particular way depending on the result of perturbation detection and the presence of new labeled and unlabeled data. The training method implements principles of self-knowledge distillation, the compactness maximization of class distribution and the interclass gap, the compression of feature representations, and consistency regularization. Consistency regularization makes it possible to utilize both labeled and unlabeled data to obtain a robust model and implement continuous adaptation. Experiments are performed on the publicly available CIFAR-10 and CIFAR-100 datasets using model backbones based on modules ResBlocks from the ResNet50 architecture and Swin transformer blocks. It is experimentally proven that the proposed prototype-based classifier head is characterized by a higher level of robustness and adaptability in comparison with the dense layer-based classifier head. It is also shown that multi-section structure and self-knowledge distillation feature conserve resources when processing simple samples under normal conditions and increase computational costs to improve the reliability of decisions when exposed to perturbations.