Deep Learning-based Intrusion Detection System using LSTM

Teuku Aldi Fadhlur Rahman, I Gede Mujiyatna, S.Kom.,M.Kom

2026 | Skripsi | ILMU KOMPUTER

Semakin canggihnya serangan siber telah mendorong pengembangan Intrusion Detection System (IDS) yang lebih tangguh dan mampu mengidentifikasi ancaman yang sudah dikenal maupun ancaman zero-day. Meskipun metode deteksi tradisional berbasis tanda tangan dan berbasis anomali memiliki keterbatasan dalam hal akurasi dan tingkat false positive, pendekatan deep learning telah menunjukkan peningkatan signifikan dalam kinerja deteksi intrusi. Penelitian ini mengusulkan Intrusion Detection System berbasis Long Short-Term Memory (LSTM) yang menggunakan dataset UNSW-NB15 untuk menyelidiki dampak kedalaman jaringan terhadap kinerja deteksi. Berbeda dengan penelitian sebelumnya yang terutama berfokus pada pembuktian kelayakan LSTM untuk deteksi intrusi, penelitian ini secara khusus mengkaji bagaimana variasi jumlah lapisan tersembunyi memengaruhi akurasi model, efisiensi komputasi, dan risiko overfitting. Lima model LSTM dengan konfigurasi yang berkisar antara satu hingga lima lapisan tersembunyi akan dikembangkan dan dilatih pada 70?taset, divalidasi pada 15%, dan diuji pada 15%, dengan 128 unit tersembunyi per lapisan dipertahankan di seluruh model untuk memastikan perbandingan yang adil. Dataset UNSW-NB15, yang berisi 2.540.044 catatan yang mencakup 9 kategori serangan dan lalu lintas normal, akan diproses terlebih dahulu melalui seleksi fitur (13 fitur utama), normalisasi Z-score, dan Synthetic Minority Over-sampling Technique (SMOTE) untuk mengatasi ketidakseimbangan kelas. Kinerja model akan dievaluasi menggunakan akurasi, presisi, recall, skor F1, efisiensi, dan waktu pelatihan untuk menentukan keseimbangan optimal antara kompleksitas model dan kemampuan deteksi. Temuan ini diharapkan memberikan wawasan mengenai hubungan antara kedalaman jaringan LSTM dan kinerja deteksi intrusi, yang berkontribusi pada optimalisasi solusi keamanan siber berbasis pembelajaran mendalam sekaligus mengatasi kekhawatiran terkait biaya komputasi dan overfitting dalam arsitektur jaringan saraf yang kompleks.

The increasing sophistication of cyber-attacks has necessitated the development of more robust intrusion detection systems (IDS) capable of identifying both known and zero-day threats. While traditional signature-based and anomaly-based detection methods have limitations in accuracy and false positive rates, deep learning approaches have demonstrated significant improvements in intrusion detection performance. This study proposes a Long Short-Term Memory (LSTM)-based intrusion detection system using the UNSW-NB15 dataset to investigate the impact of network depth on detection performance. Unlike previous research that has primarily focused on demonstrating the viability of LSTM for intrusion detection, this study specifically examines how varying the number of hidden layers affects model accuracy, computational efficiency, and the risk of overfitting. Five LSTM models with configurations ranging from one to five hidden layers will be developed and trained on 70% of the dataset, validated on 15%, and tested on 15%, with 128 hidden units per layer maintained across all models to ensure fair comparison. The UNSW-NB15 dataset, containing 2,540,044 records across 9 attack categories and normal traffic, will be preprocessed through feature selection (13 key features), Z-score normalization, and Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance. Model performance will be evaluated using accuracy, precision, recall, F1-score, efficiency, and training time to determine the optimal balance between model complexity and detection capability. The findings are expected to provide insights into the relationship between LSTM network depth and intrusion detection performance, contributing to the optimization of deep learning-based cybersecurity solutions while addressing concerns about computational cost and overfitting in complex neural network architectures.

Kata Kunci : Intrusion Detection System, Deep Learning, LSTM, UNSW-NB15, Network Security, Hidden Layers, Cyber-attack