Physical adversarial attack for fooling the Faster R-CNN object detector
Switch branches/tags
Nothing to show
Clone or download
Latest commit 8de4ec8 Sep 2, 2018
Failed to load latest commit information.
data initial commit Apr 13, 2018
imgs add link to the arxiv paper and add some figures Apr 17, 2018 Update Sep 3, 2018
robust_physical_attack.ipynb initial commit Apr 13, 2018

Robust Physical Adversarial Attack on Faster R-CNN Object Detector


This is the code repository for the ECML-PKDD 2018 paper: ShapeShifter: Robust Physical Adversarial Attack on Faster R-CNN Object Detector

The arXiv version is available at

The code included here reproduces our techniques presented in the paper.

In this work, we tackle the more challenging problem of crafting physical adversarial perturbations to fool image-based object detectors like Faster R-CNN. Attacking an object detector is more difficult than attacking an image classifier, as it needs to mislead the classification results in multiple bounding boxes with different scales. Our approach can generate perturbed stop signs that are consistently mis-detected by Faster R-CNN as other objects, posing a potential threat to autonomous vehicles and other safety-critical computer vision systems.

Install Dependencies

This repository depends on Tensorflow Object Detection API. Follow the installation instructions at

How to Run the Code

Run the ipython notebook by the command

jupyter notebook robust_physical_attack.ipynb

You can also run the code directly using this Colaboratory link. No need to download or install anything!

Videos of Targeted and Untargted Attacks

High-confidence Person Perturbation:


Transferability Experiments:

High-confidence Sports Ball Perturbation:


Transferability Experiments:

High-confidence Untargeted Attack:


Transferability Experiments:

drive_by_snapshots Snapshots of the drive-by test results. In (a), the person perturbation was detected 38% of the frames as a person and only once as a stop sign. The perturbation in (b) was detected 11% of the time as a sports ball and never as a stop sign. The untargeted perturbation in (c) was never detected as a stop sign or anything else.


Name Affiliation
Shang-Tse Chen Georgia Institute of Technology
Cory Cornelius Intel Corporation
Jason Martin Intel Corporation
Polo Chau Georgia Institute of Technology