Verfahren, Vorrichtung und Computerlesbares Medium zur Erkennung und Beseitigung von Gegnerischen Angriffen

Details

Veröffentlichungs-Nr.

EP4749532

Anmeldetag

17. September 2025

Veröffentlichung

27. Mai 2026

Rechtsraum

EP

Abstract

Embodiments of the present disclosure relate to methods for detection and removal of adversarial attacks. A computer program, a computer-readable data carrier and an apparatus are also disclosed. The method for detection of adversarial attacks comprises: obtaining a sample embedding from a layer of a machine learning model; obtaining one or more predetermined adversarial perturbation embeddings corresponding to the layer of the machine learning model; obtaining an attribution score between the sample embedding and each of the predetermined adversarial perturbation embeddings; determining that at least one similarity scores of the obtained similarity scores exceed a predetermined threshold; classifying the predetermined adversarial perturbation embeddings, corresponding to the determined at least one similarity scores, as relevant perturbation embeddings; and providing an indication that an adversarial attack is detected.

Anmelder

Firma

AUMOVIO Germany GmbH

Land

🇩🇪 Deutschland