Falsification Tool for Assessing the Robustness of End-to-End Object Detection Systems on Autonomous Vehicles

Fool deep learning models to recognize nonsense noises as traffic signs

Attacking End-to-End Object Detection Systems

• Adversarial Detection in ROS
  • Overview
  • Quick Start
    • Step 0: Prerequisites
    • Step 1: Setup the TurtleBot
    • Step 2: Setup the server
    • Step 3: Setup the browser
  • Training the model (Optional)
    • Step 1: Collect the Data
    • Step 2: Train the model
      • Training darknet yolov3-tiny
      • Training darknet yolov4-tiny
      • Training yolov4 mobilenet lite

Overview

Generating adversarial patch is as easy as drag and drop.

Quick Start

Step 0: Prerequisites

$ sudo apt install ros-noetic-desktop
$ sudo apt install ros-noetic-rosbridge-suite ros-noetic-turtlebot3-simulations ros-noetic-turtlebot3-gazebo ros-noetic-teleop-twist-keyboard

Step 1: Setup the TurtleBot

$ cd ros_ws
$ rosdep install --from-paths src --ignore-src -r -y

# For ROS, please make sure you use the system python3, rather than python from anaconda
# Deactivate conda and rm -rf build devel should solve the problem

$ catkin_make
$ source devel/setup.sh
$ export TURTLEBOT3_MODEL=waffle
$ roslaunch turtlebot3_lane turtlebot3_lane_traffic_signs.launch

Step 2: Setup the server

Pre-trained models are available here, replace the weights folder with:

https://github.com/wuhanstudio/adversarial-ros-detection/releases/tag/v1.0.0

$ roslaunch turtlebot3_lane rosbridge_websocket.launch
$ cd model
$ conda env create -f environment.yml
$ conda activate adversarial-ros-detection

# You may need to put the turtlebot on track first
# rosrun teleop_twist_keyboard teleop_twist_keyboard.py

# For Gazebo Simulator
$ python3 detect.py --env gazebo --model weights/keras/yolov4/yolov4_mobilenet_lite_3_gazebo.h5

# For real turtlebot3
$ python3 detect.py --env turtlebot --model weights/keras/yolov4/yolov4_mobilenet_lite_3_gazebo_turtlebot.h5

Optional (test models only without attacks):

# For real turblebot3
$ python3 detect_cv.py --env turtlebot --cfg weights/darknet/yolov3/yolov3-tiny-traffic-3.cfg --weights weights/darknet/yolov3/yolov3-tiny-traffic-3_turtlebot.weights --classes weights/classes.txt

# For Gazebo Simulator
$ python3 detect_cv.py --env gazebo --cfg weights/darknet/yolov3/yolov3-tiny-traffic-3.cfg --weights weights/darknet/yolov3/yolov3-tiny-traffic-3_gazebo.weights --classes weights/classes.txt

Step 3: Setup the browser

This is just a website, your can use any web server, just serve all the content under client/web.

The client is built as a single executable file.

$ ./client

For Linux and Mac, or other Unix, the server can be built with:

$ go get -u github.com/gobuffalo/packr/packr
$ go get github.com/gobuffalo/[email protected]
$ packr build

The web page will be available at: http://localhost:3333/

That's it!

Training the model (Optional)

Step 1: Collect the Data

The following script collects image data from the topic /camera/rgb/image_raw and corresponding control command in /cmd_vel. The log file is saved in driving_log.csv, and images are saved in IMG/ folder

$ cd model/data
$ # Collect center camera data
$ python3 line_follow.py --camera center --env gazebo
$ python3 ros_collect_data.py --camera center --env gazebo

Step 2: Train the model

Once the data is collected, we need to label images before training the model. This tool labelimg is used for labelling.

Training darknet yolov3-tiny

After labelling, you can use the following script to generate the training and testing set.

$ python3 train_test_split.py

Now we are ready to train the model using the darknet framework:

$ git submodule init
$ git submodule update

$ cd model/utils/darknet
$ # If you have a GPU, you may need to change the Makefile
$ # Compiling the darknet binary
$ make

$ cp -r ../darknet-conf/one-class/* ./
$ ./darknet detector train data/obj.data cfg/yolov3-tiny-traffic-3.cfg yolov3-tiny.conv.11

Finally, we can convert the darknet *.weights file to keras *.h5 file:

$ cd model/utils/keras-YOLOv3-model-set
$ python3 tools/model_converter/convert.py cfg/yolov3-tiny-traffic-3.cfg weights/yolov3-tiny-traffic-3.weights weights/yolov3-tiny-traffic-3.h5

Training darknet yolov4-tiny

$ ./darknet detector train data/obj.data cfg/yolov4-tiny-traffic-3.cfg yolov4-tiny.conv.29

Similarly, we can convert the darknet *.weights file to keras *.h5 file:

$ cd model/utils/keras-YOLOv3-model-set
$ python3 tools/model_converter/convert.py cfg/yolov4-tiny-traffic-3.cfg weights/yolov4-tiny-traffic-3.weights weights/yolov4-tiny-traffic-3.h5

Training yolov4 mobilenet lite

Before training the model, we need to convert the format of annotation files:

From:
class_id x_center y_center width height
0 0.625000 0.346875 0.081250 0.143750

To:
image_file_path x_min,y_min,x_max,y_max,class_id
path/to/img2.jpg 120,300,250,600,2

This can be done using the following script that generates a trainval.txt file:

$ cd model/utils/keras-YOLOv3-model-set/
$ cp -r ../../data/IMG ../../data/*.txt ./
$ cp ../keras-conf/* ./
$ python3 darknet_to_keras_trainval.py

Finally, we can train the model:

$ python3 train.py --model_image_size=160x320 --model_type=yolo4_mobilenet_lite --anchors_path=configs/yolo4_anchors.txt --annotation_file=trainval.txt --classes_path=obj.names --eval_online --save_eval_checkpoint

Save the model as keras:

python3 yolo.py --model_type=yolo4_mobilenet_lite --weights_path=logs/000/trained_final.h5 --anchors_path=configs/yolo4_anchors.txt --classes_path=obj.names --model_image_size=160x320 --dump_model --output_model_file=yolov4_mobilenet_lite.h5

Test the model:

python3 yolo.py --model_type=yolo4_mobilenet_lite --weights_path=yolov4_mobilenet_lite.h5 --anchors_path=configs/yolo4_anchors.txt --classes_path=obj.names --model_image_size=160x320 --image