ML Security PapersThe Luna Bound Propagator for Formal Analysis of Neural Networks
Published on arXiv
2603.23878
Input Manipulation Attack
OWASP ML Top 10 — ML01
Key Finding
Achieves competitive bound tightness and computational efficiency with state-of-the-art alpha-CROWN implementations on VNN-COMP 2025
Luna
Novel technique introduced
The parameterized CROWN analysis, a.k.a., alpha-CROWN, has emerged as a practically successful bound propagation method for neural network verification. However, existing implementations of alpha-CROWN are limited to Python, which complicates integration into existing DNN verifiers and long-term production-level systems. We introduce Luna, a new bound propagator implemented in C++. Luna supports Interval Bound Propagation, the CROWN analysis, and the alpha-CROWN analysis over a general computational graph. We describe the architecture of Luna and show that it is competitive with the state-of-the-art alpha-CROWN implementation in terms of both bound tightness and computational efficiency on benchmarks from VNN-COMP 2025.
Key Contributions
- C++ implementation of CROWN and alpha-CROWN bound propagation for production-level neural network verification
- Competitive performance with state-of-the-art Python implementations on VNN-COMP 2025 benchmarks
- General computational graph support enabling integration into existing DNN verifiers
🛡️ Threat Analysis
Bound propagation methods like CROWN and alpha-CROWN are used to formally verify neural network robustness against adversarial perturbations — this is a certified defense technique against input manipulation attacks.