Simulation-Based Verification of Neural Network Models

EasyChair Preprint 15064

Abstract

Neural networks (NNs) are increasingly being deployed in safety-critical applications such as autonomous vehicles, healthcare diagnostics, and robotics, where failures can have significant consequences. Verifying the behavior of these models is essential, yet traditional verification methods are often inadequate due to the complexity, non-linearity, and black-box nature of NNs. Simulation-based verification offers a practical alternative by testing models under a wide range of simulated scenarios to assess their robustness, safety, and reliability. This paper reviews the challenges inherent in verifying neural networks and outlines the key methods used in simulation-based approaches, such as input space sampling, adversarial testing, and robustness metrics. Additionally, we explore several case studies that demonstrate the effectiveness of simulations in uncovering potential failure modes in real-world applications. Despite its advantages, simulation-based verification has limitations, including high computational cost and incomplete coverage of the input space. To address these issues, hybrid approaches that combine simulation with formal methods are gaining traction. This paper also discusses the future directions of the field, including the need for scalable solutions, the integration of explainable AI, and the development of more sophisticated adversarial testing frameworks.Neural networks (NNs) are increasingly being deployed in safety-critical applications such as autonomous vehicles, healthcare diagnostics, and robotics, where failures can have significant consequences. Verifying the behavior of these models is essential, yet traditional verification methods are often inadequate due to the complexity, non-linearity, and black-box nature of NNs. Simulation-based verification offers a practical alternative by testing models under a wide range of simulated scenarios to assess their robustness, safety, and reliability.

Keyphrases: formal verification, neural networks, simulation-based verification, verification

@booklet{EasyChair:15064,
  author    = {Edwin Frank and Godwin Olaoye},
  title     = {Simulation-Based Verification of Neural Network Models},
  howpublished = {EasyChair Preprint 15064},
  year      = {EasyChair, 2024}}