cupbearer is a Python library for mechanistic anomaly detection. Its main purpose is to make it easy to implement either a new mechanistic anomaly detection task, or a new detection method, and evaluate it against a suite of existing tasks or methods. To that end, the library provides:

• A clearly defined interface for tasks and detectors to ensure compatibility
• Scripts and other helpers for training and evaluating detectors that satisfy this interface (to reduce the amount of boilerplate code you need to write)
• Implementations of several tasks and detectors as baselines (currently not that many)

Contributions of new tasks or detectors are very welcome! See the developer guide to get started.

The easy way: inside a virtual environment with Python >= 3.10, run

pip install git+https://github.com/ejnnr/cupbearer.git

Alternatively, if you're going to do development work on the library itself:

1. Clone this git repository
2. Create a virtual environment with Python 3.10 and activate it
3. Run pip install -e . inside the git repo to install the package in editable mode

On Linux, installing cupbearer will install pytorch with CUDA support even though this isn't necessary (we're only using pytorch for its dataloader and datasets). If you want to avoid that, run

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

before installing cupbearer.

Similarly, you may want to install jax and jaxlib manually before installing cupbearer to make sure you get the right CUDA (or CPU only) version.

You can run any script with the -h flag to see a list of options and possible values.

Let's look at an example, where we use a simple Mahalanobis distance-based detection methods to detect backdoors. First, we need to train a base model on poisoned data:

python -m cupbearer.scripts.train_classifier \
       --train_data backdoor
       --train_data.original mnist --train_data.backdoor corner \
       --model mlp \
       --dir.full logs/classifier

This will train an MLP on MNIST, where the backdoor trigger is that the top left pixel is set to white (and the model is trained to classify images with this trigger as zeros). There are many more options like learning rate etc. you can set, but the defaults should work well enough.

This script will have saved the final model to logs/classifier/model/. We can now train a Mahalanobis-based detector on this model:

python -m cupbearer.scripts.train_detector \
       --detector mahalanobis --task backdoor \
       --task.backdoor corner --task.run_path logs/classifier \
       --dir.full logs/detector

(The fact that we need to specify the backdoor trigger again is an artifact of how things are currently implemented and will likely be changed---the detector is only actually trained on clean images). train_detector loads the model from task.run_path and also uses that to figure out which dataset to use (in this case, MNIST again).

Finally, we can evaluate the detector on a mix of clean and poisoned images:

python -m cupbearer.scripts.eval_detector
       --detector from_run --detector.path logs/detector \
       --task backdoor --task.backdoor corner \
       --task.run_path logs/classifier \
       --dir.full logs/detector

Note that we reuse logs/detector as the output directory. This will add the evaluation results to the directory of the detector training run (it won't overwrite anything).

The reason we need to specify some things, such as the task, multiple times is that these scripts also support much more flexible setups. For example, we could evaluate the detector on images with a different backdoor trigger (as an ablation to check that these are not flagged as anomalous), in which case we might also want to save the results to a different directory:

python -m cupbearer.scripts.eval_detector
       --detector from_run --detector.path logs/detector \
       --task backdoor --task.backdoor noise --task.backdoor.std 0.2 \
       --task.run_path logs/classifier \
       --dir.full logs/detector_ablation \

In the future, there might be easier ways of getting "reasonable defaults" for some settings, for now you could just write small wrapper scripts for these kinds of pipelines.

Just like a cupbearer tastes wine to avoid poisoning the king, mechanistic anomaly detection methods taste new inputs to check whether they're poisoned. (Or more generally, anomalous in terms of how they affect the king ... I mean, model. I admit the analogy is becoming a bit strained here.)