Static Vulnerability Analysis of Neural Network Computational Graphs

EasyChair Preprint 16025

4 pages•Date: August 12, 2026

Abstract

We present CGVulnScan, a compiler-based static analysis frame- work for detecting security vulnerabilities in neural network com-putational graphs before deployment. Unlike existing tools that target a single vulnerability type, CGVulnScan analyzes adversarial susceptibility, numerical precision errors, and training-data leak- age together, using dataflow analysis and abstract interpretation.We introduce a vulnerability propagation model that captures how architectural patterns amplify risk through layer composition. The analysis operates directly on ONNX and TensorFlow GraphDef intermediate representations, without training data or model execution. In a case study of three representative production architectures (ResNet-50, MobileNetV2, EfficientNetV2-S), the efficiency optimized models were flagged with substantially more adversarial and privacy vulnerabilities than the standard residual baseline—up to 23× and 8.4× more in raw counts, a gap we examine (and qualify) in Section 4. CGVulnScan completes analysis in roughly 2.6 seconds per model, making it practical for CI/CD integration.

Keyphrases: Neural Network Security, Privacy Leakage, abstract interpretation, adversarial robustness, computational graphs, dataflow analysis, static analysis, vulnerability detection

Links:

https://easychair.org/publications/preprint/XwNh

BibTeX entry

BibTeX does not have the right entry for preprints. This is a hack for producing the correct reference:

@booklet{EasyChair:16025,
  author    = {Harishankar P P},
  title     = {Static Vulnerability Analysis of Neural Network Computational Graphs},
  howpublished = {EasyChair Preprint 16025},
  year      = {EasyChair, 2026}}

Download PDF Open PDF in browser