mumoRAG-attacks

Adversarial and poisoning attacks against multimodal retrieval-augmented generation (RAG)

Instructions

• Install uv: pip install uv
• Install the project as editable: uv pip install -e .
  • To install optional dependencies: uv pip install -e ".[nb,opt,test,dev]" Alternatively:
  • pip install -r requirements.txt or pip install -r requirements.opt.txt

Running

Training

To train the attack:

python src/attack_train.py --config-name <name>

Experiments are defined in the experiments module.

Fields of the config can be overridden, e.g.:

python src/attack_train.py --config-name <name> train.n_gradient_steps=10

Evaluation

To train the attack:

python src/attack_eval.py --config-name <name>

Experiment output

The experiment outputs are contained in the data directory and are split into subdirectories as follows.

Attacks

Contains the output adversarial images. File names contain the name of the experiment that generated them and a hash of the experiment parameters.

Embeddings

Contains the cached embeddings of datasets using embedding models. File names contain the name of the dataset and embedding model.

Results

Contains the results json files produced from the evals. File names contain the name of the experiment that generated them and a hash of the experiment parameters.

Sharing data

By default, files are output into the draft subdirectories of attacks and results, which are ignored. To share work, move files into the parent directories. You can then adjust the paths to point to the parent directories, either in the experiment configuration:

configstore.store(
    ...,
    node=ExperimentConfig(
        train=ExperimentTrainConfig(
            ...,
            save_folder=ATTACKS_FOLDER.parent,
        ),
        test=ExperimentEvalConfig(
            ...,
            results_folder=RESULTS_FOLDER.parent,
        ),
    )
)

or via the CLI

python src/attack_train.py --config <config_name> train.save_folder="data/attacks" eval.results_folder="data/results"