CS 6431 Final Project (Evaluating the Risk of Efficient Model Extraction Attack to Steal the Knowledge from Public OCR APIs)

Data

Data processing and annotation scripts are located under /data/script. We are undecided whether we will release our Google OCR annotation due to the ethical concerns and potential the term of use violation.

cd /data
mkdir TextOCR
mkdir TextOCR/basic
cd TextOCR/basic
wget https://dl.fbaipublicfiles.com/textvqa/data/textocr/TextOCR_0.1_train.json
wget https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip
unzip https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip

Examples of Google OCR Annotations

Models

Please follow the astep below to clone the model repo. We will release our modification patches once we resolve the conflicts.

mkdir models
cd models
git clone https://github.com/clovaai/deep-text-recognition-benchmark.git

Visualization of Validation Curves Enabled by https://wandb.ai

Trained on Human Annotations

Trained on Google OCR Annotations