Toward Scalable Verification for Safety-Critical Deep Networks

TL;DR

This paper discusses advancing scalable verification methods for deep neural networks used in safety-critical systems, aiming to improve safety guarantees by developing new techniques and design strategies.

Contribution

It introduces two approaches: scalable verification techniques and design choices that enhance the verifiability of deep learning systems.

Findings

Proposed methods improve verification scalability.

Identified design practices for more verifiable neural networks.

Framework for integrating verification into system design.

Abstract

The increasing use of deep neural networks for safety-critical applications, such as autonomous driving and flight control, raises concerns about their safety and reliability. Formal verification can address these concerns by guaranteeing that a deep learning system operates as intended, but the state of the art is limited to small systems. In this work-in-progress report we give an overview of our work on mitigating this difficulty, by pursuing two complementary directions: devising scalable verification techniques, and identifying design choices that result in deep learning systems that are more amenable to verification.