|
--- |
|
task_categories: |
|
- feature-extraction |
|
- image-classification |
|
- zero-shot-image-classification |
|
pretty_name: Perturbed Faces |
|
size_categories: |
|
- 1K<n<10K |
|
--- |
|
|
|
# Perturbed Faces |
|
|
|
This dataset contains 1000 images from [CelebA dataset](!https://www.kaggle.com/datasets/jessicali9530/celeba-dataset). For each of the thousand images dataset also has [LowKey](https://openreview.net/forum?id=hJmtwocEqzc) perturbed version and [Fawkes](https://sandlab.cs.uchicago.edu/fawkes/) perturbed version. |
|
LowKey and Fawkes perturbed images have _attacked & _cloaked at the end of the filename respectively. |
|
|
|
|
|
| File Name | Version | |
|
|---------------------|--------------------------| |
|
| 000001.jpg | Original | |
|
| 000001_cloaked.png | Fawkes perturbed version | |
|
| 000001_attacked.png | LowKey perturbed version | |
|
|
|
|
|
The Fawkes perturbed images are created using CLI provided in the [github repository](https://github.com/Shawn-Shan/fawkes) with protection mode set to mid. The LowKey version of |
|
images are created using Python code provided with the paper. |
|
|
