🛡️ CyberSentinel: Deep Learning vs. The Bad Guys

🧐 What is this?

Look, the internet is a scary place. You’ve got malware, phishing, denial-of-service attacks... basically a buffet of digital headaches. Traditional security measures are like bringing a knife to a laser fight. They just can't keep up.

This project uses Deep Learning to look at the BETH dataset (real-world logs, very fancy) and figure out who is a regular user and who is trying to burn the server room down. We are trying to predict the sus_label. If it's 1, it's malicious. If it's 0, we're chill.

💾 The Data (BETH Dataset)

We are using the BETH dataset. It's pre-processed (thank the data gods), so we don't have to clean up messy CSVs. Here is what we are looking at:

Column	Description	Type
processId	Who are you? (Process ID)	int64
threadId	Which thread is screaming?	int64
parentProcessId	Who's your daddy? (Parent Process ID)	int64
userId	ID of the user. Probably Dave from accounting.	int64
mountNamespace	Mounting restrictions.	int64
argsNum	How many arguments?	int64
returnValue	What happened? (Usually 0)	int64
sus_label	The Target. 1 = 😈, 0 = 😇	int64

🧠 The Brain (Model Architecture)

We built a simple but brave Multi-Layer Perceptron (MLP) using PyTorch.

• Input Layer: Whatever size the features are.
• Hidden Layer 1: 128 Neurons + ReLU (to keep things non-linear).
• Hidden Layer 2: 64 Neurons + ReLU.
• Output Layer: 1 Neuron + Sigmoid (because we need a probability between 0 and 1).

Tech Stack:

• Python 🐍
• PyTorch 🔥
• Pandas 🐼
• Scikit-Learn 🧠

📊 The Results (Here be dragons)

Okay, so here is the output we got. I need you to look at this closely.

• Training Accuracy: 0.11 (11%)
• Validation Accuracy: 0.06 (6%)
• Testing Accuracy: 0.92 (92%)

Wait, what? Yes, you read that right. The model performed terribly on the training data (worse than a coin flip—it's almost like it was trying to be wrong), but then absolutely crushed it on the test set with 92% accuracy.

My Theory: Either this model is a genius that learns by failing, or the loss function (CrossEntropyLoss combined with Sigmoid) is doing some weird dimensional gymnastics. But hey, 92% on the test set? We take those wins. 🚀

💻 How to Run This

If you want to witness this mathematical anomaly yourself:

1. Clone this repo.
2. Install dependencies:

   pip install torch pandas scikit-learn torchmetrics

3. Run the script:

   python train_model.py

📝 A Note on the Code

The code uses CrossEntropyLoss with a Sigmoid layer. Typically, PyTorch prefers BCEWithLogitsLoss for binary classification to avoid numerical instability, but we like to live dangerously here.