Powered by OpenAIRE graph
Found an issue? Give us feedback
addClaim

Adversarial examples in machine learning

Authors: Schuhmann, Jonas;

Adversarial examples in machine learning

Abstract

Diese Arbeit untersucht die mathematischen Grundlagen zur Generierung sogenannter Adversarial Examples und entwickelt Methoden zum Schutz von Modellen des maschinellen Lernens gegen solche gezielten Störungen. Das Konzept, bekannt als Adversarial Risk, beschreibt die Robustheit eines Modells als Min-Max-Optimierungsproblem. Adversarial Training wird als primäre Abwehrstrategie analysiert und experimentell bewertet. Um die Robustheit weiter zu verbessern, wird eine Schranke für das Adversarial Risk eingeführt. Basierend auf diesen theoretischen Erkenntnissen wird die TRADES-Methode als eine optimierte Variante des Adversarial Trainings vorgestellt. Darüber hinaus wird eine neue Klasse von Adversarial Examples eingeführt, die als Regular Adversarial Examples bezeichnet wird. Es wird ein Algorithmus entwickelt und experimentell getestet, der die Abwesenheit dieser Regular Adversarial Examples innerhalb eines bestimmten Störbereichs für neuronale Netzwerke bestätigt. Die Ergebnisse zeigen, dass dieser Ansatz moderate Robustheitsgarantien bietet, insbesondere für einfachere Netzwerkarchitekturen.

This work explores the mathematical foundations for generating adversarial examples and develops methods to safeguard machine learning models against such targeted perturbations. Central to this is the concept of adversarial risk, which allows the robustness of a model to be formulated as a Min-Max problem. Adversarial training is analyzed and experimentally evaluated as the primary defense strategy. To further enhance robustness, a bound for adversarial risk is introduced. Based on these theoretical insights, the TRADES method is presented as an optimized variant of adversarial training. Additionally, we introduce a new class of adversarial examples, referred to as regular adversarial examples. We also develop and experimentally test an algorithm designed to certify the absence of these regular adversarial examples within a specified perturbation range for neural networks. The results demonstrate that this approach provides moderate robustness guarantees, particularly for simpler network architectures.

  • BIP!
    Impact byBIP!
    selected citations
    These citations are derived from selected sources.
    This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    0
    popularity
    This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
    Average
    influence
    This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    Average
    impulse
    This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
    Average
Powered by OpenAIRE graph
Found an issue? Give us feedback
selected citations
These citations are derived from selected sources.
This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Citations provided by BIP!
popularity
This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
BIP!Popularity provided by BIP!
influence
This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Influence provided by BIP!
impulse
This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
BIP!Impulse provided by BIP!
0
Average
Average
Average
Upload OA version
Are you the author of this publication? Upload your Open Access version to Zenodo!
It’s fast and easy, just two clicks!