Awesome

Zero Gravity is a system for proving an inference pass (i.e. a classification) for a pre-trained, public Weightless Neural Network run on a private input. Zero Gravity builds upon the recent BTHOWeN model by Susskind et al (2022), in which the authors improve upon earlier WNN models in a number of interesting ways. Most importantly for this hackathon project, they helpfully provide an implementation complete with pre-trained models and reproducible benchmarks.

See our blog post for an extensive description!

Built as part of the ZKHack hackathon, Lisbon, 2023.

Setup

Clone and install the custom aleo compiler supporting lookup arguments

Usage

Values in the example below were generated using our fork of BTHOWeN.

python3 scripts/generate_aleo_code.py 56
../aleo-setup/aleo/target/debug/aleo run main "$(cat input_file.txt)" "$(cat hash_values.txt)" "$(cat bloom_filters.txt)" $(cat winning_discriminator_value.txt) "$(cat winning_discriminator_index.txt)"

Implementation notes and limitations

• All inputs are private, except for which discriminator (category) succeeded, as well as the model itself.
• Proving time of e.g. a full MNIST example is not practical yet. PRs and suggestions for improvement are welcome.
• ❌ Print statements during the build phase are expected - and will be removed in future iterations of Aleo.
• The number of lookup constraints are not included in the final print statement but will show up when compiling in debug mode.
• Note that even when changing certain inputs, the proof might succeed, as it essentially checks which discriminator won.