@ericyocam6Professor, California Polytechnic State University
Shipped: Neural Network Model Reverse Engineering Toolkit
An AI model analysis proof of concept that examines trained machine learning models to reveal architectural details, parameter counts, computational complexity, estimated memory requirements, and deployment characteristics.
Supports models from frameworks including TensorFlow/Keras, PyTorch, ONNX, and scikit-learn. The project also explores preliminary AI security considerations such as model transparency, potential attack exposure, and deployment risks.
Designed for AI developers, researchers, and security practitioners interested in understanding model internals and assessing model characteristics before deployment.
AI SecurityMachine LearningDeep LearningNeural NetworksModel AnalysisModel Reverse EngineeringAI TransparencyModel Auditing
Comments2
Post a sample of the report your toolkit produces on one well-known free AI download, especially the section on attack exposure and deployment risks. A security person deciding whether to let an outside AI into their company would understand what you built from that one page much faster than from the list of supported tools.
Great suggestion, Ryan! I agree that showing an actual model analysis report would make the toolkit's value much clearer, particularly for enterprise security teams evaluating third-party AI models.
Using a publicly available model would provide a practical example of its architecture, parameter counts, computational requirements, and potential deployment risks. I'd also want to distinguish between risks identified through static model analysis and vulnerabilities that require adversarial testing to confirm.
That distinction is important because understanding a model's architecture is only the first step toward evaluating its security. A sample report with clearly documented findings and limitations would be a useful way to demonstrate what the toolkit can and cannot assess. Appreciate the suggestion!
2 comments
Post a sample of the report your toolkit produces on one well-known free AI download, especially the section on attack exposure and deployment risks. A security person deciding whether to let an outside AI into their company would understand what you built from that one page much faster than from the list of supported tools.
Great suggestion, Ryan! I agree that showing an actual model analysis report would make the toolkit's value much clearer, particularly for enterprise security teams evaluating third-party AI models.
Using a publicly available model would provide a practical example of its architecture, parameter counts, computational requirements, and potential deployment risks. I'd also want to distinguish between risks identified through static model analysis and vulnerabilities that require adversarial testing to confirm.
That distinction is important because understanding a model's architecture is only the first step toward evaluating its security. A sample report with clearly documented findings and limitations would be a useful way to demonstrate what the toolkit can and cannot assess. Appreciate the suggestion!