ModelScan

Import ModelScan reports in JSON format. ModelScan inspects serialized machine-learning models (pickle, TensorFlow SavedModel, Keras, NumPy) for operators that are able to execute code when the model is loaded.

Generate a report with:

modelscan -p /path/to/models -r json -o modelscan.json

Severity Mapping#

ModelScan assigns a severity to each unsafe operator it knows about, and DefectDojo maps that scale directly:

ModelScan reports an unsafe operator rather than a known vulnerability, so its findings carry no CVE or CWE. The operator is recorded in vuln_id_from_tool as module.operator.

Sample Scan Data#

Sample ModelScan scans can be found here.

Default Deduplication Hashcode Fields#

By default, DefectDojo identifies duplicate Findings using these hashcode fields:

• title
• file_path
• vuln_id_from_tool