Research map / Adversarial robustness and security / Certified and empirical robustness

Sound Verification of Deployed Neural Networks

NeurIPS 2026 · Adversarial robustness and security › Certified and empirical robustness · approach: Neural network robustness verification

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Addresses soundness of neural network verifiers when applied to the deployed (floating-point) version of a network, rather than its idealized real-valued version.

Problem

Verifiers are supposed to be sound (never certify a property that fails), but recent work shows that known verifiers claimed sound are not guaranteed sound for the deployed floating-point network.

Approach

Abstract text indicates the paper aims to build verifiers that are guaranteed sound for the deployed version of a neural network; the specific technique is not stated in the abstract.

Surprising findings

• Per the abstract, none of the currently known verifiers claimed to be sound are guaranteed to be sound for the deployed version of a network.

Closest papers by meaning

• No Soundness in the Real World: On the Challenges of the Verification of Deployed Neural Networks (ICML 2025 · similarity 0.93)
• Stress-Testing Neural Network Verifiers with Provably Robust Instances (NeurIPS 2026 · similarity 0.86)
• Clip-and-Verify: Linear Constraint-Driven Domain Clipping for Accelerating Neural Network Verification (NeurIPS 2025 · similarity 0.85)
• Dynamic Back-Substitution in Bound-Propagation-Based Neural Network Verification (AAAI 2025 · similarity 0.83)
• SDP-CROWN: Efficient Bound Propagation for Neural Network Verification with Tightness of Semidefinite Programming (ICML 2025 · similarity 0.83)
• Training Verification-Friendly Neural Networks via Neuron Behavior Consistency (AAAI 2025 · similarity 0.83)