search

melihcatal / advsecurenet

Machine Learning Security Library
https://melihcatal.github.io/advsecurenet/
MIT License
4 stars 3 forks source link

CLI Attacks doesn't support custom number of classes and input channels #119

Closed melihcatal closed 8 months ago

melihcatal commented 10 months ago

The default value for num_input_channels is set to 3 in the model_factory class https://github.com/melihcatal/advsecurenet/blob/aed5b78caf85fcede73d900f5f4fd108d9552be1/advsecurenet/models/model_factory.py#L45

However, the attack configuration files, such as the YML file for FGSM attack

https://github.com/melihcatal/advsecurenet/blob/aed5b78caf85fcede73d900f5f4fd108d9552be1/advsecurenet/configs/cli/attacks/fgsm_attack_config.yml#L3-L6

do not provide a parameter to modify this value. This leads to size mismatches and consequent failures when executing attacks on MNIST models through the CLI.