PRADANA, Fadli Dony and Farikhin, Farikhin and Warsito, Budi (2026) PENGEMBANGAN MODEL STACKING HYBRID LIGHTGBM, RANDOM FOREST DAN META-LEARNER MULTI LAYER PERCEPTRON UNTUK KLASIFIKASI SERANGAN WEB. Masters thesis, UNIVERSITAS DIPONEGORO.

	Text cover awal.pdf Download (169kB)
	Text Cover.pdf Restricted to Repository staff only Download (2MB)
	Text BAB I.pdf Download (285kB)
	Text BAB II.pdf Download (790kB)
	Text BAB III.pdf Restricted to Repository staff only Download (787kB)
	Text BAB IV.pdf Restricted to Repository staff only Download (2MB)
	Text BAB V.pdf Restricted to Repository staff only Download (275kB)
	Text Daftar Pustaka.pdf Download (312kB)
	Text Lampiran.pdf Restricted to Repository staff only Download (1MB)

Abstract

Serangan siber pada aplikasi web kian kompleks dan sering luput dari IDS berbasis signature, khususnya SQL Injection, XSS, dan Brute Force. Penelitian ini membangun sistem klasifikasi serangan web dengan pendekatan stacking hybrid yang menggabungkan LightGBM dan Random Forest sebagai base-learner serta Multi Layer Perceptron (MLP) sebagai meta-learner. Tahap pra pemrosesan data meliputi pembersihan data, winsorization, label encoding, normalisasi berbasis data latih, seleksi fitur menggunakan ANOVA F-Test, serta penerapan SMOTE pada data latih. Penelitian ini menggunakan dataset CIC-IDS2017 subset Web Attack, dengan evaluasi kinerja model berdasarkan metrik akurasi, precision, recall, dan F1-score pada skenario klasifikasi biner dan multikelas. Hasil menunjukkan akurasi 99,4% dan Macro-F1 0,902 pada skenario biner serta akurasi 99% dan Macro-F1 0,67 pada skenario multiclass, dengan keunggulan signifikan atas baseline terkuat tervalidasi melalui Bootstrap Paired T-Test per sampel pada Macro-F1 dengan CI95 0,116 hingga 0,226 dan p<0.0001. Analisis SHAP menegaskan peran fitur perilaku lalu lintas dan dinamika antar paket, selaras dengan daftar ANOVA Top 15. Uji di VPS menunjukkan latensi 0,000013 detik dan memori efisien, serta generalisasi pada dataset UNSW NB15, MTA KDD 19, dan KDD Cup 99. Dengan demikian, penelitian ini menyajikan kerangka klasifikasi serangan web yang akurat dan efisien, sehingga dapat menjadi landasan pengembangan sistem deteksi intrusi berbasis pembelajaran mesin di masa mendatang.
Kata Kunci : Stacking Hybrid, LightGBM, Random Forest, Multi Layer Perceptron, Klasifikasi Serangan Web

Cyberattacks on web applications are becoming increasingly complex and often evade signature-based IDS, particularly in the case of SQL Injection, XSS, and Brute Force attacks. This study proposes a web-attack classification framework using a stacking-hybrid architecture that integrates LightGBM and Random Forest as base learners, with a Multi-Layer Perceptron (MLP) serving as the meta-learner.. The data preprocessing stage includes data cleaning, winsorization, label encoding, training-data-based normalization, feature selection using the ANOVA F-test, and the application of SMOTE to the training data. Experiments are conducted on the CIC-IDS2017 Web Attack subset under binary and multiclass schemes, evaluated using accuracy, precision, recall, and F1-score. The results show an accuracy of 99.4% and a Macro-F1 of 0.902 for the binary scenario, and an accuracy of 99% and a Macro-F1 of 0.67 for the multiclass scenario, with a statistically significant improvement over the strongest baseline as validated by a sample-wise bootstrap paired t-test on the Macro-F1, yielding a 95% confidence interval of 0.116 to 0.226 and p < 0.0001. SHAP analysis confirms the importance of traffic-behavior features and inter-packet dynamics, which are consistent with the ANOVA Top-15 feature list. Experiments on a VPS demonstrate a latency of 0.000013 seconds and efficient memory usage, as well as good generalization on the UNSW-NB15, MTA-KDD’19, and KDD Cup 99 datasets. Accordingly, this study presents an accurate and efficient web attack classification framework that can serve as a foundation for the development of future machine-learning-based intrusion detection systems.
Keywords : Stacking Hybrid, LightGBM, Random Forest, Multi Layer Perceptron, Web Attack Classification

Actions (login required)

View Item