🥄There is no spoon

There is no spoon is an application to generate adversarial images.

Quick Start

With Docker [recommended]:

• pre-requisite: You need to have docker installed. You can find the documentation here https://docs.docker.com/install/
• put the image you want to change in your current directory in png format. Feel free to use sample_image.png in this project
• run the following command and change MY_IMAGE.PNG with the name of your image:

 sudo docker pull amagash/there_is_no_spoon:latest
 sudo docker run -v $(pwd):/data amagash/there_is_no_spoon --input /data/MY_IMAGE.PNG

• the output image will be saved in your current directory under the name "adversarial_image.png".

With pip:

• pre-requisite: I would recommend you to create a virtual environment first:

virtualenv -p python3.7 spoon_env
source spoon_env/bin/activate
pip install there_is_no_spoon 'h5py<3.0.0'

• put the image you want to change in your current directory in png format. Feel free to use sample_image.png in this project
• run the following command and change MY_IMAGE.PNG with the name of your image:

there_is_no_spoon --input MY_IMAGE.PNG --output adversarial_image.png

• the output image will be saved in your current directory under the name "adversarial_image.png".

Advanced features

Several arguments can be passed to the docker to customize your adversarial image generation.

Synopsis

sudo docker run -v $(pwd):/data amagash/there_is_no_spoon --input /data/MY_IMAGE.PNG
[--output=OUTPUT_PATH][--mode=MODE][--target_class=TARGET_CLASS][--target_score=TARGET_SCORE]
[--learning_rate=LEARNING_RATE][--max_change=MAX_CHANGE]

Flags

--input=INPUT_PATH
This sets your file input path. Make sure to use an image in png format.

--output=OUTPUT_PATH
This sets your file destination output. By default it is the current directory with docker. If you use pip, you need to specify it. Make sure to precise a png format for the image output.

--mode=MODE
You can run the code in 2 modes: "predict" or "generate".

• The "predict" mode simply takes an image as an input and prints the classification of this image using InceptionV3 as a model.
• The "generate" mode let's you generate an adversarial image from an input image of your choice. By default, mode="generate".

--target_class=TARGET_CLASS
The class index corresponding to the new object you would like to change your image to. For instance, if you have a picture of a cat and you want to generate and adversarial image from the cat image that will be classified as being a spoon, you need to find the corresponding index for the spoon classification. You can find the different class and their index here: https://storage.googleapis.com/download.tensorflow.org/data/imagenet_class_index.json

By default, the target_class=910, corresponding to the classification of a wooden spoon (because there is no spoon 😉).

--target_score=TARGET_SCORE
The target score is the minimum score you would like to reach for the new classification (which should be between 0 and 1). For instance, if you have a picture of a cat and you want to generate and adversarial image from the cat image that will be classified as being a spoon with a confidence of at least 95%, you need to set your target_score to 0.95. By default, target_score = 0.98 to reach at least 98% confidence for the new classification.

--learning_rate=LEARNING_RATE
The learning rate corresponds to how much to update the adversarial image in each iteration. Lower value will make your model more precise but might also slow the generation of the adversarial image. By default, learning_rate=0.5.

--max_change=MAX_CHANGE
The maximum change each pixel can support. By default, max_change=0.1. Larger number produces an image faster but risks more distortions.