CICIDS2017

• hat tip to @r-dube for best learning rate/batch size and insights re: data prep and neural network architecture.

cicids2017_best_run.ipynb (97.7837% val accuracy)

• Val accuracy 97.7837 at epoch 18 (ran for 35, but think 20 epochs better for performance).
• Standardized features, 1 hidden layer (26 neurons)
• Linear, ReLU, BatchNorm1d, Linear
• Cross Entropy Cost Function
• lr=0.004, decaying at 0.5 every 10 epochs
• bs=8

cicids2017_2020_07_23_97_val_acc.ipynb (97.1394% val accuracy)

• Val accuracy 97.1394% at epoch 19
• Standardized features, 1 hidden layer (26 neurons)
• Linear, Sigmoid, BatchNorm1d, Linear, Sigmoid,
• Cross Entropy Cost function
• lr=0.004, decaying at 0.5 every 10 epochs
• bs=8

cicids2017_2020_07_23.ipynb

• Best run: validation accuracy 80%
• Fully connected neural network with one hidden layer (26 hidden neurons)
• Both activation functions are sigmoid
• Cross entropy cost function
• Best run: learning_rate=0.0001, batch_size=500, weight_decay=0, epochs: ~250

cicids2017_2020_07_23_96.ipynb

• Best run: validation accuracy 96.8%
• Used MinMaxScaler(), but incorrectly because validation data was included in the scaling. Need to rescale using only training data then transform val data based on training data.