The DriveByFlyBy dataset is a first of its kind, designed to evaluate the effectiveness of backdoor attacks against object detectors in real-world scenarios.
The dataset was created in response to the lack of public datasets featuring physical object triggers deployed under challenging real-world conditions, particularly for tasks like traffic sign and vehicle detection.
It contains more than 40 diverse scenarios of physical object trigger deployments in the wild with four different kinds of triggers and two detection tasks. This dataset fills a critical gap by providing a resource to assess the robustness of object detection models against physical backdoor attacks in real-world settings. It allows researchers to explore the impact of environmental factors on attack effectiveness, advancing our understanding of attacks and the development of defense mechanisms.
This dataset is released from our ACSAC Paper. To demonstrate attacks in the wild using our dataset, we also created a website to showcase a number of videos for research purposes.
This github repository contains the code for attacks validated with our dataset, scripts and documentation required to run our pipeline.
If you use this dataset, please cite us as follows
Bao Gia Doan, Dang Quang Nguyen, Callum Lindquist, Paul Montague, Tamas Abraham, Olivier De Vel, Seyit Camtepe, Salil S. Kanhere, Ehsan Abbasnejad, and Damith C. Ranasinghe. On the Credibility of Backdoor Attacks Against Object Detectors in the Physical World. In Annual Computer Security Applications Conference (ACSAC), 2024.
This dataset is made available for academic use only. However, we take your privacy seriously! If you find yourself or personal items in this dataset and feel unhappy about their use, please contact us via this email (drivebyflyby@gmail.com) and we will immediately remove the respective data from the dataset.