A Hybrid Defense Method against Adversarial Attacks on Traffic Sign Classifiers in Autonomous Vehicles

TL;DR

This paper proposes a hybrid defense approach combining transfer learning, random filtering, ensembling, and local feature mapping to improve traffic sign classification robustness against adversarial attacks in autonomous vehicles.

Contribution

It introduces a novel hybrid defense method that enhances DNN resilience to adversarial attacks on traffic sign classifiers using multiple strategies.

Findings

Achieves 99% accuracy without attacks

Maintains 88% accuracy under various adversarial attacks

Outperforms traditional defense methods by up to 55% in accuracy

Abstract

Adversarial attacks can make deep neural network (DNN) models predict incorrect output labels, such as misclassified traffic signs, for autonomous vehicle (AV) perception modules. Resilience against adversarial attacks can help AVs navigate safely on the road by avoiding misclassication of signs or objects. This DNN-based study develops a resilient traffic sign classifier for AVs that uses a hybrid defense method. We use transfer learning to retrain the Inception-V3 and Resnet-152 models as traffic sign classifiers. This method also utilizes a combination of three different strategies: random filtering, ensembling, and local feature mapping. We use the random cropping and resizing technique for random filtering, plurality voting as ensembling strategy and an optical character recognition model as a local feature mapper. This DNN-based hybrid defense method has been tested for the no attack scenario and against well-known untargeted adversarial attacks (e.g., Projected Gradient Descent or PGD, Fast Gradient Sign Method or FGSM, Momentum Iterative Method or MIM attack, and Carlini and Wagner or C&W). We find that our hybrid defense method achieves 99% average traffic sign classification accuracy for the no attack scenario and 88% average traffic sign classification accuracy for all attack scenarios. Moreover, the hybrid defense method, presented in this study, improves the accuracy for traffic sign classification compared to the traditional defense methods (i.e., JPEG filtering, feature squeezing, binary filtering, and random filtering) up to 6%, 50%, and 55% for FGSM, MIM, and PGD attacks, respectively.