Is there any struct which stores the parameter values set for training the neural network architecture? I am interested to access specifically the model parameters in the code. Where can I find them?
Any particular reason to retrieve training parameters on your device? We store metadata about DSP parameters into the model_metadata.h file but not for the training parameters.
You can retrieve training parameters using our API if it’s something that works for you: https://docs.edgeimpulse.com/reference/edge-impulse-api/learn/keras-settings
Thanks @aurel for your reply!
I am trying to secure the model generated from the platform on the edge using hardware security. The simple idea behind being, no intruder has access to model parameters in order to tweak them or change the output.
For example- I have a model running on edge to classify gestures, and based on the classified outputs I have an associated actuation. According to me, there would be three things that makes the model unique enough to generate outputs in the way they do:
- Feature engineering parameters
- Model variables
- Model itself (in case its an IP)
To me then it makes sense to store 1. and 2. in hardware security modules and in the run-time retrieve them post some validation. I already figured out feature engineering parameters but seems like model variables are not stored the same way and hence the question.
Also, your inputs would be valuable to me in understanding if it’s right to secure model variables.
Thanks in advance!