Certainty Attacks Using Explainability Preprocessing

Carina Newen

doi:10.1007/978-3-032-02215-8_15

Certainty Attacks Using Explainability Preprocessing

2026-01-01

0

PRE

AI

C

Carina Newen *

S

Sofia Vergara Puccini

E

Emmanuel Müller

DOI：10.1007/978-3-032-02215-8_15

Original

Original request for help

Save

Abstract

En 中文

With the importance of machine learning rising over the years, learners have been perfected according to their performance on unseen data without considering that the data to be classified could be actively manipulated by an adversary to cause misinterpretations. In this work, we propose to attack the certainty of models. We consider this an important attack angle, as a lot of countermeasures for detection rely on certainty metrics. The new aspect of this paper is that we optimized our attack on four key aspects: The success rate, confidence in the misclassification, transferability of attacks to other models, and the image quality of the generated adversarial. We are introducing this as a means to improve attacks in general regarding key aspects such as certainty of the attack and other desirable attack metrics that do not limit themselves to accuracy. The code can be found at https://github.com/KDDOpenSource/Certainty- Attacks.git.

Keywords:

Certainty Attacks

Adversarial Examples

Explainability Preprocessing

Model Robustness

Transferability

Certainty Attacks Using Explainability Preprocessing

Abstract

Journal

Organization