README.md

Laboratory 1 for the AIC Course

This repository contains all the materials, scripts, and documentation for Laboratory 1 of the AI and Cybersecurity course.

Overview

This lab builds a complete intrusion detection pipeline on a curated subset of the CICIDS2017 dataset using Feed Forward Neural Networks (FFNN) in PyTorch.

The laboratory is structured into six progressive tasks that comprehensively cover the intrusion detection pipeline:

• Task 1: Data cleaning, stratified splits, outlier inspection, scaling comparison (Standard vs Robust).
• Task 2: Shallow FFNN (single hidden layer) with neuron sweep and activation (Linear vs ReLU).
• Task 3: Feature bias analysis (Destination Port), port substitution experiment, feature removal impact.
• Task 4: Class imbalance mitigation via class‐weighted CrossEntropy.
• Task 5: Deep architectures, batch size impact, optimizer comparison (SGD / Momentum / AdamW).
• Task 6: Overfitting and regularization (Dropout, BatchNorm, Weight Decay) on deeper models.

Repository Structure

Laboratory1/
├── lab/            # Data, notebooks and support material
├── report/         # LaTeX source files for the lab report
├── resources/      # Additional resources (e.g., links, PDFs, images)
└── README.md       # This file

Note

The detailed lab report, including all experimental results and analysis, can be found here.

Lab Objectives & Requirements

Objectives:

1. Understand preprocessing choices (scaling, outlier retention).
2. Evaluate architectural depth vs minority class detection.
3. Quantify bias induced by a single feature (Destination Port).
4. Mitigate class imbalance using weighted loss.
5. Compare optimizers and batch sizes for convergence/generalization.
6. Assess regularization techniques on tabular intrusion data.

Requirements:

• Python 3.10+
• PyTorch, scikit-learn, numpy, pandas, matplotlib, seaborn
• Dataset file: lab/data/dataset_lab_1.csv

Quick Start

1. Clone:

   git clone <repo_url>
   cd Laboratory1
   

2. Create environment (example with venv):

   python -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt
   

   (Create requirements.txt if missing; minimal list: torch torchvision torchaudio scikit-learn pandas numpy seaborn matplotlib)

3. Run notebook:

   jupyter notebook lab/notebooks/Lab1_FFNN.ipynb
   

4. Results (plots, metrics) saved under lab/results/images/<task>_plots/.

Data

Place dataset_lab_1.csv in lab/data/.
No automatic download is performed (course-provided subset).

Reproducing Experiments

• Set random seed (already fixed to 42 in notebook).
• To switch scaler: change X_train_use = X_train_std to robust variant.
• To rerun port bias test: execute Task 3 cells after initial training.

Results Summary (High-level)

• Best shallow (ReLU, 64 neurons) balanced macro F1.
• Deep 3-layer [32,16,8] + AdamW gave strong trade-off.
• Weight decay (1e-4) sufficed; heavy Dropout/BatchNorm harmed minority recall.
• Port feature induced spurious correlation—removal reduced PortScan shortcuts.

Authors

Name	GitHub	LinkedIn	Email
Andrea Botticella
Elia Innocenti
Simone Romano