Harnessing Generative Modeling and Autoencoders Against Adversarial Threats in Autonomous Vehicles

The safety and security of Autonomous Vehicles (AVs) have been an active area of interest and study in recent years. To enable human behavior, Deep Learning (DL) and Machine Learning (ML) models are extensively used to make accurate decisions. However, the DL and ML models are susceptible to various...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on consumer electronics Jg. 70; H. 3; S. 6216 - 6223
Hauptverfasser: Raja, Kathiroli, Theerthagiri, Sudhakar, Swaminathan, Sriram Venkataraman, Suresh, Sivassri, Raja, Gunasekaran
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York IEEE 01.08.2024
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Schlagworte:
Adversarial attacks
Autonomous vehicles
Closed box
Decisions
Deep learning
generative denoising autoencoders
Glass box
Machine learning
neural structured learning
Noise
Noise reduction
Perturbation methods
Training
Vehicles
ISSN:0098-3063, 1558-4127
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The safety and security of Autonomous Vehicles (AVs) have been an active area of interest and study in recent years. To enable human behavior, Deep Learning (DL) and Machine Learning (ML) models are extensively used to make accurate decisions. However, the DL and ML models are susceptible to various attacks, like adversarial attacks, leading to miscalculated decisions. Existing solutions defend against adversarial attacks proactively or reactively. To improve the defense methodologies, we propose a novel hybrid Defense Strategy for Autonomous Vehicles against Adversarial Attacks (DSAA), incorporating both reactive and proactive measures with adversarial training with Neural Structured Learning (NSL) and a generative denoising autoencoder to remove the adversarial perturbations. In addition, a randomized channel that adds calculated noise to the model parameter is utilized to encounter white-box and black-box attacks. The experimental results demonstrate that the proposed DSAA effectively mitigates proactive and reactive attacks compared to other existing defense methods, showcasing its performance by achieving an average accuracy of 80.15%.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0098-3063
1558-4127
DOI:10.1109/TCE.2024.3437419