Scalable Parallel Verification of Quantized Neural Networks via MIQCP Optimization Encoding
IEEE Access · 2026 · 14 · 117026-117045
DOI: 10.1109/access.2026.3718887 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shifu Yang und 1 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: project of China National Petroleum Corporation, titled “Research on Key Technologies for Fracture Identification and Characterization of Carbonate Reservoirs”.
Abstract (Originalsprache)
Quantized Neural Networks are widely used in safety-critical systems due to their efficiency, yet they remain vulnerable to adversarial perturbations. Existing verification approaches based on the Big- $M$ Mixed-Integer Linear Programming (MILP) suffer from loose relaxations and high computational cost. To address these issues, we propose a compact Mixed-Integer Quadratic Constrained Programming (MIQCP) encoding that introduces quadratic constraints to tighten feasible regions and reduce spurious branching. We further integrate DeepPoly-based symbolic interval analysis, which refines variable bounds and reduces the number of quadratic constraints. To enhance scalability, we design a sub-model decomposition and parallel verification framework, enabling efficient verification of large networks. We implement these techniques in PIQV, a parallel verifier for integer quantized networks. Experiments show that PIQV substantially outperforms a controlled Big- $M$ MILP baseline (EQV*) built on the same abstract domain, resolving more instances in 11 of 12 large-perturbation settings (91.7%) with an average runtime reduction of 49.7% and up to a $12\times $ speedup on deeper networks; the advantage further generalizes from fully connected to convolutional INT8 models and from MNIST to Fashion-MNIST.
Förderung
- project of China National Petroleum Corporation, titled “Research on Key Technologies for Fracture Identification and Characterization of Carbonate Reservoirs”
Zitieren
APA
Yang, S., & Yang, W. (2026). Scalable Parallel Verification of Quantized Neural Networks via MIQCP Optimization Encoding. IEEE Access, 14, 117026-117045. https://doi.org/10.1109/access.2026.3718887
Vancouver
Yang S, Yang W. Scalable Parallel Verification of Quantized Neural Networks via MIQCP Optimization Encoding. IEEE Access. 2026;14:117026-117045. doi:10.1109/access.2026.3718887
BibTeX
@article{Yang2026,
author = {Yang, Shifu and Yang, Wuyang},
title = {Scalable Parallel Verification of Quantized Neural Networks via MIQCP Optimization Encoding},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {117026--117045},
doi = {10.1109/access.2026.3718887}
}
EUVIMED