How protectai/modelscan works
ModelScan is a static scanner for unsafe code in serialized ML models. A user supplies a model through the CLI; the scan orchestrator prepares the model and dispatches format-specific checks, which report findings, errors, or skipped scans for presentation. The README documents Pickle, H5, Keras, and TensorFlow SavedModel support, with severity-ranked findings. The source index supports the main module relationships shown; because several implementations are only partially excerpted, some wiring is omitted rather than inferred.
Main entry points: modelscan.py Scan orchestrator · cli.py Command-line interface · picklescanner.py Pickle opcode analysis · model.py Model stream and context
Components 18
User entry
- Command-line interface
modelscan/cli.py - CLI utilities
modelscan/tools/cli_utils.py
Scan runtime
- Scan orchestrator
modelscan/modelscan.py - Scan settings
modelscan/settings.py - Middleware pipeline
modelscan/middlewares/middleware.py - Extension format detection
modelscan/middlewares/format_via_extension.py - Model stream and context
modelscan/model.py - Scanner interface
modelscan/scanners/scan.py
Format scanners
- Pickle and PyTorch scans
modelscan/scanners/pickle/scan.py - Pickle opcode analysis
modelscan/tools/picklescanner.py - H5 Lambda scan
modelscan/scanners/h5/scan.py - Keras archive scan
modelscan/scanners/keras/scan.py - SavedModel scans
modelscan/scanners/saved_model/scan.py
Findings and output
- Issue model and severity
modelscan/issues.py - Scan errors
modelscan/error.py - Skip records
modelscan/skip.py - Report output
modelscan/reports.py
Other components
- Model provider
Connections 23
- Model provider → Command-line interface: supplies model
README.md - Command-line interface → Scan orchestrator: starts scan
modelscan/cli.py - Command-line interface → Scan settings: loads settings
modelscan/cli.py - Command-line interface → CLI utilities: uses utilities
modelscan/cli.py - Scan orchestrator → Scan settings: uses settings
modelscan/modelscan.py - Scan orchestrator → Middleware pipeline: runs middleware
modelscan/modelscan.py - Middleware pipeline → Extension format detection: configures middleware
modelscan/middlewares/format_via_extension.py - Extension format detection → Model stream and context: sets format context
modelscan/middlewares/format_via_extension.py - Scan orchestrator → Model stream and context: prepares model
modelscan/modelscan.py - Scan orchestrator → Scanner interface: uses scanner interface
modelscan/modelscan.py - Scan orchestrator → Scan errors: handles errors
modelscan/modelscan.py - Scan orchestrator → Skip records: handles skips
modelscan/modelscan.py - Scan orchestrator → Issue model and severity: collects findings
modelscan/modelscan.py - Pickle and PyTorch scans → Pickle opcode analysis: analyzes opcodes
modelscan/scanners/pickle/scan.py - Pickle and PyTorch scans → Model stream and context: reads model
modelscan/scanners/pickle/scan.py - H5 Lambda scan → SavedModel scans: reuses operator checks
modelscan/scanners/h5/scan.py - Keras archive scan → SavedModel scans: reuses operator checks
modelscan/scanners/keras/scan.py - SavedModel scans → Model stream and context: reads model
modelscan/scanners/saved_model/scan.py - Pickle opcode analysis → Issue model and severity: creates findings
modelscan/tools/picklescanner.py - Pickle opcode analysis → Scan errors: raises scan errors
modelscan/tools/picklescanner.py - Pickle opcode analysis → Skip records: records skips
modelscan/tools/picklescanner.py - Report output → Scan orchestrator: formats scan results
modelscan/reports.py - Report output → Issue model and severity: renders findings
modelscan/reports.py