search

inoueke-n / optimization-detector

Optimization detection over compiled binaries
MIT License
7 stars 2 forks source link

It is about the paper #1

Closed anzosasuke closed 3 years ago

anzosasuke commented 3 years ago

I could not understand padding section of the paper. Why did you truncate and padding? Can you may be explain?

anzosasuke commented 3 years ago

why pre-padding increases accuracy?

davidepi commented 3 years ago

In our study the input vector has a fixed size of 2048 bytes. Usually the binary file is bigger than that amount and there is no problem. However, we need to use padding during training (training only!), so the network can learn how to classify when we have less than 2048 bytes available.

If at some point, during testing or inference, the amount of data available is less than 2048 bytes, we can just fill the remaining with 0. The network already learned how to deal with that during training.

So pre-padding does not increase accuracy, but prevents accuracy drops if we use less bytes during testing or inference.

anzosasuke commented 3 years ago

thank you so much.