LAROUDIE, M. Clément (2024) Investigate Certifiability using Morphological Operators PRE - Research Project, ENSTA.

[img]PDF
Available under License Creative Commons Attribution.

881Kb

Abstract

Deep Neural Networks are vulnerable to adversarial attacks which become problematic in safety-critical applications that require guarantees. One of the challenges is therefore to create models that are less susceptible to attack, but which can also be certified. Our work seeks to investigate a particular morphological operator, the dilation, which has several interesting properties and could be an interesting alternative to convolutions. To this end, we attempted to build a 1-Lipschitz model which incorporates dilation layers.

Item Type:Thesis (PRE - Research Project)
Uncontrolled Keywords:Deep Learning, Lipschitz models, Robustness, Certifiability, Morphology, Computer vision
Subjects:Information and Communication Sciences and Technologies
ID Code:9990
Deposited By:Clément LAROUDIE
Deposited On:13 août 2024 15:45
Dernière modification:13 août 2024 15:45

Repository Staff Only: item control page