Adversarial attack algorithm for traffic sign recognition

Deep learning suffers from the threat of adversarial attacks, and its defense methods have become a research hotspot. In all applications of deep learning, intelligent driving is an important and promising one, facing serious threat of adversarial attack in the meanwhile. To address the adversarial...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Multimedia tools and applications Ročník 84; číslo 29; s. 35137 - 35149
Hlavní autoři: Wang, Juan, Shi, Lei, Zhao, Yang, Zhang, Haoxi, Szczerbicki, Edward
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York Springer US 01.09.2025
Springer Nature B.V
Témata:
1231: IoT-driven Computer Vision Technology for Smart Transportation Applications
Algorithms
Artificial intelligence
Black boxes
Computer Communication Networks
Computer Science
Data Structures and Information Theory
Deep learning
Efficiency
Multimedia Information Systems
Neural networks
Object recognition
Signs
Speaking
Special Purpose and Application-Based Systems
Street signs
Traffic control
Traffic signs
Voice recognition
Black box
Traffic sign recognition
Adversarial attack
Algorithm security
ISSN:1573-7721, 1380-7501, 1573-7721
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Deep learning suffers from the threat of adversarial attacks, and its defense methods have become a research hotspot. In all applications of deep learning, intelligent driving is an important and promising one, facing serious threat of adversarial attack in the meanwhile. To address the adversarial attack, this paper takes the traffic sign recognition as a typical object, for it is the core function of intelligent driving. Considering that the black box attack does not need to know the internal characteristics of the model, it can have more practical value. However, the existing black box attack algorithm has high visit time and low efficiency in attacking sample generation. In this regard, the SimBA algorithm with high efficiency is selected and improved according to the characteristics of traffic signs, named the L-SimBA algorithm. According to the graphic characteristics of traffic signs that are already known, L-SimBA algorithm limits the search subspace consciously and specifies the set of search directions, and that is the core idea of it. By this way, L-SimBA algorithm can generate adversarial samples faster. Experimental comparison shows that in the field of traffic sign recognition, L-SimBA algorithm is better than SimBA algorithm. On the premise of obtaining similar quality adversarial attack samples, the success rate of adversarial measures gets higher, and the number of model visits reduces considerably, thus the attack efficiency of the algorithm improves greatly.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1573-7721
1380-7501
1573-7721
DOI:10.1007/s11042-022-14067-5