Detecting offending images

This repository provides the sources code for the publication: TODO

The required datasets can be downloaded here:

SMID: https://osf.io/2rqad/

ImageNet: https://image-net.org/download.php

Reproducing the results

The evaluation of ImageNet-based pre-trained models as well as the prompt optimization

The experiments can be executed by running the bash script ./main/run_files/run_cvs.sh

Detecting offending images contained in ImageNet1k and 21k

The bash script ./main/run_files/run_eval_imagenet.sh can be used to reproduce the papers experiments. The underlying scripts are contained in main/check_datasets and can be adapted to any Dataset which follows the pytorch torchvision.datasets.ImageFolder class.

Notebooks to reproduce figures

Further the optimized prompts to detect offending images with CLIP models are provided in /results .

The list of as possible offensive detected images are provided in /results .

The notebooks contained in /main/notebooks can be used to reproduce the paper's figures.