On the (In)Security of Loading Machine Learning Models - 2026 IEEE Symposium on Security and Privacy - Artifacts

Read full text Summarise with VIRAT
Authors: Gabriele Digregorio, Marco Di Gennaro, Stefano Zanero, Stefano Longari, Michele Carminati
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

πŸ† Distinguished Artifact Award πŸ† - IEEE Symposium on Security and Privacy 2026 This repository contains the full artifact package for the paper: β€œOn the (In)Security of Loading Machine Learning Models” (2026 IEEE Symposium on Security and Privacy). The package is organized to support artifact evaluation along three dimensions: Availability: raw data, scripts, notebooks, PoCs, and model artifacts are included. Reproducibility: results can be recomputed from the provided artifacts (survey statistics/tests, plots, and PoC executions). Functionality: scripts and PoCs run as intended and produce the expected outputs. Mapping between folders and paper sections/results 1) Vulnerability PoCs Folder: vulnerabilities/ Paper mapping: - KV1, KV2, KV3 β†’ Section 4.1 - SV1, SV2, SV3 β†’ Section 4.2 Goal: PoCs achieve arbitrary code execution at model load time, despite framework-level security measures (e.g., Keras safe_mode), with success indicated by spawning /bin/sh during loading. Each vulnerability subfolder includes: a README.md with instructions, a report.md snapshot, a docker/ environment, and the PoC artifacts/scripts. To run and verify all six PoCs automatically: cd vulnerabilities/ python3 run.py 2) Hugging Face scanning experiments Folder: HF_experiments/ Paper mapping: - Section 4.3, Table 3 Goal: availability of all PoC artifacts used for the Hugging Face tests. 3) Survey analysis Folder: survey/ Paper mapping: - Section 5 (UP2) Goal: the provided raw responses and analysis artifacts reproduce the reported survey statistics, plots, and Wilcoxon perception-shift results. Contains: - raw survey CSV, - survey form copy, - analysis/plot scripts, - notebook versions, - Wilcoxon perception-shift test notebook, - docker/ environment for reproducible execution. 4) Keras version adoption study Folder: version_adoption_keras/ Paper mapping: - Appendix B, Figure 2 Goal: the provided query output and plotting artifacts regenerate the same Keras version-adoption trend shown in Figure 2. Contains: - BigQuery SQL query, - raw CSV export, - script and notebook to regenerate the plot, - docker/ environment for reproducible execution. Contacts For questions, clarifications, or collaboration inquiries: Gabriele Digregorio β€” gabriele.digregorio@polimi.it Marco Di Gennaro β€” marco.digennaro@polimi.it Stefano Zanero β€” stefano.zanero@polimi.it Stefano Longari β€” stefano.longari@polimi.it Michele Carminati β€” michele.carminati@polimi.it Updates This repository contains a snapshot from the artifact evaluation phase. Updates are tracked in the GitHub mirror: https://github.com/necst/security-model-sharing

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Security and Verification in Computing Web Application Security Vulnerabilities

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN