ModelScan

What is ModelScan?

ModelScan is an open-source security tool developed by Protect AI that scans machine learning models to detect unsafe code and protect against model serialization attacks. As the industry's first ML model scanner supporting multiple formats, it examines model files without executing them, identifying malicious code injected during the serialization process. The tool protects users from credential theft, data theft, data poisoning, and model poisoning when using ML models from various sources. ModelScan supports PyTorch, TensorFlow, Keras, sklearn, and XGBoost frameworks, scanning H5, Pickle, and SavedModel formats, and integrates seamlessly into ML and CI/CD pipelines.

Key Features

• Multi-Format Support

  Scans models across multiple serialization formats including H5, Pickle, and SavedModel, protecting users working with PyTorch, TensorFlow, Keras, sklearn, and XGBoost.

• Static Analysis Security Scanning

  Examines model files byte-by-byte without executing code, detecting unsafe operations and malicious code signatures safely and efficiently.

• Risk Severity Classification

  Categorizes detected vulnerabilities into four severity levels: CRITICAL, HIGH, MEDIUM, and LOW, enabling prioritized remediation.

• CI/CD Pipeline Integration

  Works seamlessly with CI/CD pipelines and ML workflows, allowing automated security scanning before model deployment, retraining, or evaluation.

• Open Source and Free

  Available as an Apache 2.0 open-source project under pip install, free for all users to secure their machine learning workflows.

Use Cases

• Pre-Deployment Security Validation : Scan all models before deploying to endpoints to ensure they haven't been compromised during storage or transmission.
• ML Supply Chain Protection : Detect supply chain attacks by scanning pre-trained models before retraining, fine-tuning, or evaluation to prevent environment contamination.
• Third-Party Model Verification : Verify safety of models shared from internet, between teams, or downloaded from Hugging Face before loading them into production systems.
• Post-Training Security Audit : Scan models after training completes to detect any malicious code injection that may have occurred during the training pipeline.
• Enterprise ML Security Compliance : Integrate automated model scanning into enterprise ML workflows to meet security compliance requirements for AI/ML software deployment.

FAQs