ML Models
Review machine learning model scanning support in JFrog Xray.
Introduction
JFrog Xray provides security analysis for machine learning (ML) models throughout the software development lifecycle. ML model scanning focuses on identifying malicious or unsafe model files and detecting known model formats for software bill of materials (SBOM) visibility. For more information about malicious model detection, see Detect Malicious AI Models.
Beyond security detection, models identified by Xray can be managed within the JFrog AI Catalog, providing a centralized registry for the discovery, curation, and policy-based approval of machine learning assets.
Capabilities
The following capabilities are available for ML models.
| Capability | HuggingFace ML | Generic Model Detection |
|---|---|---|
| SBOM and Model Identification |  | |
| CVE Detection | Not applicable | Not applicable |
| License Detection | |  |
| Malicious Model Detection | | |
| Operational Risk | Not applicable | Not applicable |
Source code scanning isn't applicable for ML models. All capabilities are available through binary scanning only.
Binary Scanning
Binary scanning analyzes ML model files stored in JFrog Artifactory repositories. Xray identifies model formats, detects malicious models, and provides SBOM visibility for ML assets.
Supported Model Types
The following ML model types are supported.
| Model Type | Supported |
|---|---|
| HuggingFace ML | |
| Generic Model Detection | |
Supported Model File Formats
Xray supports these ML model file formats: bin, ckpt, dill, flax, ggml, gguf, h5, hdf5, joblib, keras, mpk, msgpack, nemo, npy, npz, onnx, pb, pdparams, pkl, pt, pth, safetensors, tflite, and zip.
Supported ML Frameworks
Xray supports these ML frameworks: Flax, GGML, GGUF, Joblib, Keras H5, NeMo, NumPy Archive, NumPy Array, ONNX, PaddlePaddle, Pickle/Dill, PyTorch Archive, PyTorch state_dict, Safetensors, SavedModel, and TFLite.
Detection Contexts
The following detection contexts are supported for ML models.
| Context | Supported |
|---|---|
| Dedicated ML model repositories | |
| Generic Artifactory repositories | |
| Inside Docker container images | |
Additional Information
- Xray identifies ML model binaries in both dedicated and Generic repositories, as well as inside Docker containers.
- Generic Model Detection provides SBOM visibility, including format identification and operational risk, but not malicious model scanning.
- HuggingFace ML provides the richest analysis with license detection and malicious model detection.
- CVE detection isn't applicable for ML models, as they don't have traditional software vulnerabilities.
Related Topics
Updated 25 days ago