Luna: A C++ Bound Propagator for Neural Network Verification
A team of researchers has unveiled Luna, a novel bound propagator based on abstract interpretation, developed in C++ for the formal evaluation of neural networks. This tool accommodates Interval Bound Propagation, DeepPoly/CROWN analysis, and alpha-CROWN analysis across a general computational graph. In benchmarks from VNN-COMP 2025, Luna surpasses the leading alpha-CROWN implementation in both computational efficiency and bound tightness. Unlike existing alpha-CROWN implementations that are limited to Python, Luna's C++ framework simplifies integration into current DNN verifiers and production systems, enhancing long-term maintenance. The research paper detailing Luna can be found on arXiv with the identifier 2603.23878. Luna is also accessible to the public on GitHub.
Key facts
- Luna is a new abstract-interpretation-based bound propagator implemented in C++.
- Luna supports Interval Bound Propagation, DeepPoly/CROWN analysis, and alpha-CROWN analysis.
- Luna operates over a general computational graph.
- Luna outperforms the state-of-the-art alpha-CROWN implementation in bound tightness and computational efficiency.
- Benchmarks from VNN-COMP 2025 were used for evaluation.
- Luna is publicly available on GitHub.
- Existing alpha-CROWN implementations are limited to Python, complicating integration.
- The paper is available on arXiv with identifier 2603.23878.
Entities
Institutions
- arXiv
- VNN-COMP