Perkembangan Internet of Things (IoT) meningkatkan konektivitas antar perangkat, namun juga memperluas risiko serangan siber, khususnya Distributed Denial of Service (DDoS). Salah satu tantangan utama dalam sistem deteksi serangan adalah tingginya tingkat false negative, yaitu serangan yang gagal terdeteksi oleh sistem. Penelitian ini bertujuan untuk menganalisis performa beberapa algoritma machine learning dalam meminimalkan false negative pada deteksi serangan DDoS di jaringan IoT berbasis gateway. Empat algoritma diuji, yaitu Random Forest, Multi-Layer Perceptron (MLP), LightGBM, dan Logistic Regression menggunakan dataset UNSW-NB15. Tahap prapemrosesan meliputi normalisasi MinMaxScaler, seleksi fitur SelectKBest (Chi-Square), serta penyeimbangan kelas menggunakan ADASYN. Hasil eksperimen menunjukkan bahwa Random Forest dan MLP memperoleh nilai AUC tertinggi sebesar 0,98, diikuti LightGBM (0,97) dan Logistic Regression (0,94). Analisis confusion matrix menunjukkan bahwa LightGBM menghasilkan false negative terendah, sementara Random Forest juga menunjukkan sensitivitas tinggi dengan performa stabil. Meskipun MLP memiliki akurasi tinggi dan mampu menangkap pola non-linear, ukuran model yang besar menjadi pertimbangan dalam implementasi pada perangkat dengan keterbatasan sumber daya. Berdasarkan hasil tersebut, LightGBM direkomendasikan sebagai model paling optimal karena mampu menyeimbangkan sensitivitas deteksi, efisiensi komputasi, dan ukuran model dalam implementasi pada IoT gateway. Abstract The rapid growth of the Internet of Things (IoT) has significantly enhanced device connectivity but has also increased vulnerability to cyber threats, particularly Distributed Denial of Service (DDoS) attacks. One of the main challenges in intrusion detection systems is the high false negative rate, where attacks remain undetected. This study aims to evaluate the effectiveness of several machine learning algorithms in minimizing false negatives for DDoS detection within gateway-based IoT networks. Four algorithms were examined: Random Forest, Multi-Layer Perceptron (MLP), LightGBM, and Logistic Regression using the UNSW-NB15 dataset. Data preprocessing included MinMaxScaler normalization, SelectKBest (Chi-Square) feature selection, and ADASYN oversampling to address class imbalance. Experimental results indicate that Random Forest and MLP achieved the highest AUC scores of 0.98, followed by LightGBM (0.97) and Logistic Regression (0.94). Confusion matrix analysis reveals that LightGBM produced the lowest false negative rate, while Random Forest also demonstrated high sensitivity and stable performance. Although MLP showed strong classification capability in handling non-linear patterns, its larger model size may limit deployment in resource-constrained environments. Overall, LightGBM is recommended as the most optimal model due to its balance between detection sensitivity, computational efficiency, and compact model size for IoT gateway implementation.