Table of Contents
Fetching ...

The Luna Bound Propagator for Formal Analysis of Neural Networks

Henry LeCates, Haoze Wu

Abstract

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.

The Luna Bound Propagator for Formal Analysis of Neural Networks

Abstract

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.
Paper Structure (27 sections, 1 equation, 2 figures, 1 table)

This paper contains 27 sections, 1 equation, 2 figures, 1 table.

Figures (2)

  • Figure 1: High-level overview of 's system architecture.
  • Figure 2: Example Python API usage: loading an ONNX model, setting input bounds, and computing $\alpha$-CROWN bounds.