跳转至

文章背景与核心概要

深度神经网络（DNNs）在视觉感知系统中得到了广泛应用，然而其对对抗攻击的脆弱性带来了重大的安全隐患，特别是在自动驾驶等安全攸关的应用中。传统的防御机制在资源受限的环境中往往面临成本效率过低的问题。

本文介绍了一种新颖、高效且有效的对抗攻击检测方案，该方案利用了复杂视觉系统中的多任务感知能力。通过分析多个并行视觉任务（如目标检测和实例分割）推理输出之间的一致性差异，该方法能够成功标记出对抗性扰动。作者开发了一种专门的一致性评分指标、一种用于选择表现最佳模型对的优化方法，并在 BDD100k 验证数据集上使用投影梯度下降（PGD）攻击验证了其防御效果，实现了高达 99.9% 的 ROC-AUC 检测性能。

Multi-Task Consistency-based Detection of Adversarial Attacks

Summary

Deep Neural Networks (DNNs) are widely used in vision perception systems, but their vulnerability to adversarial attacks poses significant safety concerns, particularly in safety-critical applications like autonomous driving. Traditional defense mechanisms are often too cost-inefficient for resource-constrained environments.

This paper introduces a novel, efficient, and effective adversarial attack detection scheme that leverages multi-task perception in complex vision systems. By analyzing inconsistencies between the inference outputs of multiple parallel vision tasks (such as object detection and instance segmentation), the method successfully flags adversarial perturbations. The authors developed a specialized consistency score metric, an optimization approach to select the best-performing model pairs, and validated their defense against Projected Gradient Descent (PGD) attacks using the BDD100k validation dataset, achieving an impressive 99.9% ROC-AUC detection performance.

Paper Metadata

Abstract

Deep Neural Networks (DNNs) have found successful deployment in numerous vision perception systems. However, their susceptibility to adversarial attacks has prompted concerns regarding their practical applications, specifically in the context of autonomous driving. Existing defenses often suffer from cost inefficiency, rendering their deployment impractical for resource-constrained applications. In this work, we propose an efficient and effective adversarial attack detection scheme leveraging the multi-task perception within a complex vision system. Adversarial perturbations are detected by the inconsistencies between the inference outputs of multiple vision tasks, e.g., object detection and instance segmentation. To this end, we developed a consistency score metric to measure the inconsistency between vision tasks. Next, we designed an approach to select the best model pairs for detecting inconsistencies effectively. Finally, we evaluated our defense against PGD attacks across multiple vision models on the BDD100k validation dataset. The experimental results demonstrated that our defense achieved a ROC-AUC performance of 99.9% detection within the considered attacker model.

Access & Resources