Giters

BishopFox / eyeballer

Convolutional neural network for analyzing pentest screenshots

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Why did you choose the neural network architecture currently being used?

c-flaherty opened this issue · comments

commented

In particular,

  1. Why does the current architecture downsample screenshots to (224, 224)? Did you find that the model is just as performant at such a low resolution?
  2. Why did you decide to use MobileNet (pre-trained on ImageNet) as opposed to other models that have performed better on ImageNet (and are also included in Keras) such as EfficientNet (any of the sizes) or Inception ResNet V2? Here is the link to the rankings for ImageNet, for convenience: https://paperswithcode.com/sota/image-classification-on-imagenet
commented

It's entirely possible that there's more optimal choices for ANY of the hyperparameters. I won't claim otherwise. :)

But in our testing, it worked really well and increasing the image size didn't seem to help. Though who knows, maybe with more (or just different) data it could be different. Definitely feel free to tinker with that.

commented

It's entirely possible that there's more optimal choices for ANY of the hyperparameters. I won't claim otherwise. :)

But in our testing, it worked really well and increasing the image size didn't seem to help. Though who knows, maybe with more (or just different) data it could be different. Definitely feel free to tinker with that.

Got it, thank you for the context!