University of Twente Student Theses

Login

Robust training using a push-pull inhibition layer for adversarial robustness in convolutional neural networks

Navarro Di Pasquale, J. (2022) Robust training using a push-pull inhibition layer for adversarial robustness in convolutional neural networks.

[img] PDF
3MB
Abstract:Adversarial attacks have gained considerable attention in recent years due to increasing real-world, safety-critical applications of Deep Neural Networks. The vulnerability against such attacks spans multiple domains and thus exhibits security concerns, mainly because they are challenging to detect, and an understanding of their existence is lacking. Consequently, the research community has proposed many defense strategies to inherently induce robustness properties through domain-specific design or training mechanisms. This paper concentrates on defending against adversarial attacks within the image classification domain, where so-called adversarial examples are constructed by carefully crafting (imperceptible) perturbations on an image such that a classifier produces erroneous predictions with high confidence. More Specifically, we quantitatively analyze an approach that builds upon a biologically-inspired component called the push-pull layer that increases robustness against naturally distorted/corrupted images. We combine the said component with adversarial training to investigate its robustness-efficacy against various adversarial attacks and threat models. The findings in this experimental study indicate that the approach allows the component to translate its properties to adversarial examples and, with further research, may prove itself as a general-purpose defense tool.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science BSc (56964)
Link to this item:https://purl.utwente.nl/essays/89463
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page