Model analysis

Model analysis

ML Model Analysis

Model Investigation

Using Keras

from tensorflow import keras
from keras.models import load_model

model = load_model("example.h5")

# Summarization
print(model.summary())

# Configuration
print(model.get_config())

# List inputs
print(model.inputs)
# List outputs
print(model.outputs)

Using PyTorch

If we don’t have torchinfo, we need to install it at first.

pip install torchinfo

Here is the code for investigation.

import torch
from torchinfo import summary

model = torch.load("example.pt")
model.eval() # it's not required for investigation only but required when inferening

batch_size = 16
print(summary(model=model, input_size=(batch_size, 3, 16, 16)))

# Also simply show model's state dict
print(model.state_dict)

Scan Model

ModelScan

ModelScan is a machine learning model scanner to protect againt Model Serialization Attacks.

# -p: Path to the file
modelscan -p example.h5
modelscan -p example.pt

# Scan all models in Hugging Face Repository
modelscan -hf owner/model-repository-name

Model Fingerprinting

Reference: Crucible

When we don't know which model is being used, we can use various approaches to identify it.

For Image Recognition Models

• Input an image that is rotated.

• Input an image that is changed its color.

• Input random images.

• Replicate a model from ImageNet because many image recognition models are based on ImageNet.

References

• PyTorch