LVBW

Official implementation of black-box label ownership verification via label backdoor watermarking (LVBW). Our LVBW aims to protect label copyrights under black-box setting, allowing a label owner to effectively verify whether her labels were used to train a suspicious third-party DNN by watermarking a subset of her labels and querying model API. Our codes is based on [https://github.com/SewoongLab/FLIP].

In this repo

It splits into three main folders: experiments, modules, and schemas. The experiments folder contains subfolders and .toml configuration files on which an experiment may be run. The modules folder stores source code for each of the subsequent part of an experiment. These modules take in specific inputs and outputs as defined by their subseqeunt .toml documentation in the schemas folder. Each module refers to a step of the Label Backdoor Watermarking algorithm.

Existing modules:

1. base_utils: Utility module, used by the base modules.
2. train_expert: Step 1 of watermarking: training expert models and recording trajectories.
3. generate_labels: Step 2 of watermarking: generating poisoned labels from trajectories.
4. select_flips: Step 3 of watermarking: strategically flipping labels within some budget.
5. train_user: Use watermarked labels to train a model. Evaluate watermarking by ACC/ASR of this model and get verification results（i.e.,p-value and $\Delta P$）.

More documentation can be found in the schemas folder.

Supported Datasets:

1. CIFAR-10
2. CIFAR-100
3. Tiny ImageNet

Installation:

conda env create -f environment.yml conda activate LVBW

Setting up:

To initialize an experiment, create a subfolder in the experiments folder with the name of your experiment:

mkdir experiments/[experiment name]

In that folder initialize a config file called config.toml. An example can be seen here: experiments/example_attack/config.toml.

The .toml file should contain references to the modules that you would like to run with each relevant field as defined by its documentation in schemas/[module name].

Running a module:

At the moment, all experiments must be manually run using:

python run_experiment.py [experiment name] [config file's name]

The experiment will automatically pick up on the configuration provided by the file.

As an example, to run the example_attack experiment one could run:

python run_experiment.py example_attack config

More module documentation can be found in the schemas folder.