EL-GY-9163: Machine Learning for Cyber Security
Implemented Pruning Defence learnt in Machine Learning for Cyber Security. The Channels in the Final Pooling Layer have been pruned in decreasing order of their average activation values across the entire validation dataset. The Models that have been saved are the repaired models when accuracy drops by at least {2%,4%,10%}.
On top of this, a GoodNet has been designed which combines the bad model and the repaired model for each of the 3 cases and we then compare the Accuracy and Attack Success Rate between the pruned models and the good nets.
The Accuracy and Attack Success Rate is not substantially different from the pruned models. It is possible that the Attacker is aware of the defender's intentions of pruning the model and using Fine-Pruning Defense and has taken measures such as training the poisoned images on a pruned model and a fine-pruned model in order to deal with the circumvent this defence strategy.
Download the zip file using the given link
https://drive.google.com/file/d/15sxgJdO9U6OkMJ4gZ80ur1_Xi1cDp1Da/view?usp=sharing
Unzip the zip file and upload each file individually on to colab with data being the root folder or Upload the unzipped folder along with its contents to Google Drive (Quicker Approach and the one used)
checkpoints folder - contains the weight for the good nets of all 3 cases
repaired_nets folder - contains the repaired models