Representasi Fitur Berbasis Deep Learning Untuk Meningkatkan Performa Nids Berbasis Machine Learning

Vian Handika, Prof. Dr. Ir. Jazi Eko Istiyanto, M.Sc., IPU, ASEAN Eng.;Prof. Dr.techn. Ahmad Ashari, M.I.Kom.

2025 | Tesis | S2 Ilmu Komputer

Kompleksitas data jaringan dan serangan siber yang berdimensi tinggi serta bersifat non-linier menjadi tantangan utama dalam pengembangan sistem deteksi intrusi (NIDS) berbasis machine learning. Representasi fitur memegang peranan penting dalam menangani tantangan ini, namun pendekatan konvensional yang umum digunakan belum sepenuhnya mampu menangkap pola non-linier yang kompleks. Di sisi lain, studi yang secara khusus mengevaluasi kontribusi pendekatan deep learning dalam membentuk representasi fitur dan dampaknya terhadap peningkatan klasifikasi model NIDS masih terbatas.

Penelitian ini mengevaluasi kontribusi tiga model representasi fitur berbasis deep learning, yaitu Sparse Autoencoder (SAE), Wide and Deep (WnD), dan Numerical Embedding (NUME), dalam meningkatkan performa klasifikasi model NIDS berbasis machine learning. Representasi laten yang dihasilkan oleh ketiga model digunakan sebagai input untuk tiga algoritma klasifikasi tradisional: Decision Tree (DTC), K-Nearest Neighbor (KNN), dan Support Vector Classifier (SVC). Evaluasi dilakukan menggunakan tiga dataset benchmark, yaitu CSE-CIC-IDS 2018, NSL-KDD, dan UNSW-NB15, mencakup proses preprocessing, transformasi fitur, pelatihan model, serta analisis performa dan visualisasi struktur data untuk menilai peningkatan separabilitas antar kelas.

Hasil penelitian menunjukkan bahwa representasi fitur berbasis deep learning (DL) secara efektif mampu meningkatkan performa deteksi model machine learning (ML), dengan peningkatan mencapai hingga 8.2%. Peningkatan tertinggi dicapai melalui representasi fitur dari model NUME, diikuti oleh WND dan SAE, yang secara konsisten mengungguli pendekatan konvensional seperti PCA dan LDA dengan selisih performa antara 1–26%. Selain itu, hasil visualisasi juga mendukung temuan ini, menunjukkan bahwa representasi fitur dari model-model DL memberikan kontribusi positif terhadap peningkatan kinerja model ML pada ketiga dataset yang digunakan. 

The complexity of network data and the high dimensional and non linear nature of cyberattacks present significant challenges in the development of machine learning based Network Intrusion Detection Systems (NIDS). Feature representation plays a critical role in addressing these challenges. However, conventional approaches commonly employed have not been fully effective in capturing complex non linear patterns. Furthermore, there is a lack of studies that specifically evaluate the contribution of deep learning approaches to feature representation and their impact on the improvement of NIDS classification models.

This study examines the contribution of three deep learning feature representation models, namely Sparse Autoencoder (SAE), Wide and Deep (WND), and Numerical Embedding (NUME), in enhancing the classification performance of machine learning based NIDS models. The latent representations generated by these models are used as input for three traditional classification algorithms: Decision Tree (DTC), K Nearest Neighbor (KNN), and Support Vector Machine (SVM). The evaluation is conducted using three benchmark datasets, which are CSE CIC IDS 2018, NSL KDD, and UNSW NB15. The evaluation process includes preprocessing, feature transformation, model training, performance analysis, and data structure visualization to assess improvements in class separability.

The results demonstrate that deep learning feature representations can effectively enhance the detection performance of machine learning models, with improvements of up to 8.2 %. The highest performance gain is achieved using the NUME model, followed by WND and SAE, all of which consistently outperform conventional methods such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) with performance differences reaching up to 26 %. In addition, the visualization results support these findings, indicating that feature representations derived from deep learning models contribute positively to the performance improvement of machine learning models across the three datasets.

Kata Kunci : Representasi Fitur, Deep Learning, Intrusion Detection System (NIDS)