REAP (REalistic Adversarial Patch) Benchmark

REAP数据集由加州大学伯克利分校创建，包含14,651个真实的交通标志图像，用于模拟对抗性补丁攻击。数据集考虑了标志的姿态、位置和光照条件，通过几何和光照变换将数字生成的补丁真实地应用于标志上。REAP数据集旨在解决自动驾驶车辆等依赖摄像头的网络物理系统中对抗性攻击的评估问题，提供了一个大规模的标准化基准，以更准确地评估新的攻击和防御措施。The REAP dataset, developed by the University of California, Berkeley, contains 14,651 real-world traffic sign images for simulating adversarial patch attacks. The dataset takes into account the pose, position, and lighting conditions of traffic signs, and realistically applies digitally generated patches to the signs via geometric and photometric transformations. The REAP dataset aims to address the evaluation of adversarial attacks in camera-dependent cyber-physical systems such as autonomous vehicles, providing a large-scale standardized benchmark for more accurate assessment of novel attack and defense measures.

搜集汇总

背景与挑战

背景概述

REAP基准是一个大规模、现实的对抗性补丁数据集，用于评估对抗性补丁攻击在真实世界条件下的有效性。它基于Mapillary Vistas数据集构建，包含超过14,000个交通标志，并通过几何和光照变换增强现实性，支持对自动驾驶等系统的安全研究。数据集提供了完整的代码和工具，方便用户进行攻击生成、模型训练和性能评估。