Smaller is Better: Enhancing Transparency in Vehicle AI Systems via Pruning

TL;DR

AI in vehicles must be transparent and trustworthy, yet post-hoc explanations for traffic-sign classifiers are often unreliable. The authors systematically compare natural training, adversarial training, and pruning, evaluating gradient-based explanations (VG, IG, SG) with sparsity (Gini) and faithfulness (ROAD) on LISA and GTSRB. They find pruning consistently yields more faithful and comprehensible saliency maps and improves efficiency, while adversarial training can increase robustness at the cost of explanation quality. Across two traffic-sign datasets and multiple architectures, pruning emerges as a practical route to transparent, resource-efficient vehicular AI with actionable guidance for deployment.

Abstract

Connected and autonomous vehicles continue to heavily rely on AI systems, where transparency and security are critical for trust and operational safety. Post-hoc explanations provide transparency to these black-box like AI models but the quality and reliability of these explanations is often questioned due to inconsistencies and lack of faithfulness in representing model decisions. This paper systematically examines the impact of three widely used training approaches, namely natural training, adversarial training, and pruning, affect the quality of post-hoc explanations for traffic sign classifiers. Through extensive empirical evaluation, we demonstrate that pruning significantly enhances the comprehensibility and faithfulness of explanations (using saliency maps). Our findings reveal that pruning not only improves model efficiency but also enforces sparsity in learned representation, leading to more interpretable and reliable decisions. Additionally, these insights suggest that pruning is a promising strategy for developing transparent deep learning models, especially in resource-constrained vehicular AI systems.

Figures (23)

• Figure 1: A figure showing heatmaps on an American traffic sign for naturally trained, adversarially trained and different pruned models for three explanation methods: Vanilla Gradient simonyan2014deepinsideconvolutionalnetworks, Integrated Gradient sundararajan2017axiomaticattributiondeepnetworks and SmoothGrad smilkov2017smoothgrad
• Figure 2: Saliency maps comparison for LISA dataset (VGG) between natural training, adversarial training and pre-train pruning.
• Figure 3: Saliency maps comparison for LISA dataset (VGG) between natural training, adversarial training and test-time pruning with no fine-tuning.
• Figure 4: Saliency maps comparison for LISA dataset (VGG) between natural training, adversarial training and test-time pruning with fine-tuning.
• Figure 5: Saliency maps comparison for LISA dataset between natural training, adversarial training ($\epsilon=0.1$) and post-train pruning (with fine-tuning) for VGG.