Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{196203,
        author = {Atharva Khambete and Avanish Kulkarni},
        title = {Secure-Torch: A Defense-in-Depth Trust Gateway for AI Model Deserialization},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {6057-6061},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=196203},
        abstract = {These instructions give you guidelines for preparing papers for the International Journal of Innovative Research in Technology (IJIRT). Model files have become an overlooked software supply-chain boundary. Common machine learning loading paths deserialize artifacts with limited trust controls, enabling attack classes such as pickle gadget execution, metadata-triggered payloads, and unsafe custom operator loading. We present secure-torch, a Python library that inserts a fail-closed trust enforcement layer into model loading workflows. The system provides a fixed-order six-stage pipeline: format detection, signature verification, threat scoring, policy enforcement, optional sandbox isolation, and controlled load return. The implementation supports PyTorch pickle-based artifacts, SafeTensors, and ONNX, combining static artifact inspection with provenance and policy gates. Evaluation against known common vulnerabilities and exposures (CVEs) indicates strong baseline coverage for deserialization attack patterns and practical compatibility with incremental adoption via audit mode. We discuss the architecture, implementation trade-offs, and limitations relevant to advancing secure ML deployment research.},
        keywords = {Model supply chain security, provenance verification, PyTorch security, safe deserialization, sandboxing, threat scoring.},
        month = {April},
        }