eleutherai / radioactive-lab

Adapting the "Radioactive Data" paper to work for text models

Radioactive data

This is EleutherAI's reimplementation of the paper "Radioactive data: tracing through training".

Their GitHub repo can be found here. Warning: The official open source implementation has a fairly complicated framework, as well as miscellaneous issues like hard-coded paths that will prevent you from running it.

We spent a fair amount of time getting the original implementation working with CIFAR10, then chose to do a full refactor after understanding the core requirements.

Our implementation is a refined version designed to demonstrate the marking, training and detection steps using the CIFAR10 dataset, with each stage self contained within its own python module. We have added logging to TensorBoard for your visualization enjoyment.

Prior to implementing the training stage we created a full working example of a resnet18 classifer trained on CIFAR10. This can be used to benchmark your particular hardware and choose a good optimizer prior to running the main code. It's also a good starting point for ML beginners.

Install

First setup pytorch. Example for GPU enabled using conda:

conda create --name radioactive_lab python=3.8
conda activate radioactive_lab
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

After that use pip to install other requirements:

pip install -r requirements.txt

If anything is missing, simply pip install [missing requirement].

Running

Please follow the basic_example.ipynb example.

For any other questions please visit our Discord!

License

This repository is licensed under the CC BY-NC 4.0.

radioactive-lab's Issues

Convert hard-coded links to relative paths (text)

The stock code contains hard-coded paths to the data. We need to make a data directory with relative paths.

Implement a Google Collab notebook for easier investigation of the code base

Applying the methodology to text

This issue is for discussing how to apply the methodology to text data, with a focus on which aspects differ from image data and why.

Refactor underlying framework

Now that we have the basic framework up and running, we want to update the code to be in accordance with best practices.

Implement CIFAR 10 on resnet18

Now that we have the basic functionality working, Laurence is going to integrate a serious deep learning model into the code as a skills building exercise and to ensure he is comfortable working with the code.

Applying the methodology to generative models

This issue is for discussing how to apply the methodology to generative models, with a focus on which aspects differ from classification models and why.

Replicate the paper's experiments.

To ensure that our refactored code is working properly, we need to replicate the experiments from the paper. We want to replicate the following experiments:

• Results (Table 2, Figure 6)
• Architecture transfer (Table 3)
• Transfer to other datasets (Table 4)
• Robustness to distillation (Table 5)

Improved detection

Instead of doing the hypothesis testing carried out in the paper, an alternative methodology for detecting radioactive data seems feasible: Take your random vector u and compute the "pure adversarial example," the random noise that the model thinks only contains the feature associated with vector u. Then send that image forward through the network. Since it's just random noise, most models should classify it randomly. However a model trained on radioactive data will see the feature u and know how to classify it.