TinyML Security Scanner

A command-line tool to analyze TensorFlow Lite models for embedded ML applications.

Features

• Metadata Extraction: Model size, architecture, input/output shapes, quantization type
• Weight Analysis: Statistical analysis of model weights with anomaly detection
• Quantization Detection: Automatically identifies float32, float16, int8, and mixed quantization

Installation

python -m venv tflite_venv
source tflite_venv/bin/activate  # On Windows: tflite_venv\Scripts\activate
pip install -r requirements.txt

Usage

Basic scan:

python src/cli.py model.tflite

Skip weight analysis (faster):

python src/cli.py model.tflite --skip-weights

Example Output

============================================================
MODEL METADATA
============================================================
Size: 0.42 MB (438,476 bytes)
Tensors: 8
Input shape: (1, 784)
Output shape: (1, 10)
Quantization: float32
Operators: MatMul

============================================================
WEIGHT ANALYSIS
============================================================
Total parameters: 109,184
Weight tensors analyzed: 3

No anomalies detected

============================================================
SUMMARY
============================================================
Status: PASS
Errors: 0
Warnings: 0

Detected Anomalies

The scanner detects:

• All zeros: Weight layer is entirely zeros (model broken)
• Constant weights: All weights identical (initialization failed)
• Extreme sparsity: >99.5% weights are zero (unusual pruning)
• High sparsity: >95% weights are zero (may be intentional)
• Unusual range: Weight values outside typical range

Testing

pytest tests/

Development

Generate test models:

python tests/generate_test_models.py

Inspect models:

python tests/inspect_models.py

Exit Codes

• 0: Success (no issues)
• 1: File not found
• 2: Invalid model format
• 10: Warnings found
• 20: Critical errors found
• 99: Unexpected error

Project Structure

src/
├── cli.py                 # Command-line interface
├── core/
│   ├── model_loader.py    # Model loading and validation
│   └── exceptions.py      # Custom exceptions
├── analyzers/
│   ├── metadata.py        # Metadata extraction
│   └── weights.py         # Weight analysis
└── reporting/
    └── text_reporter.py   # Output formatting

License

MIT