Multimodal Learning Menggunakan Efficientnetv2 Dan Distilbert Untuk Deteksi Website Ilegal Dan Implementasinya Pada Browser Extension

Sahal Maghfud (Program Studi Sistem Informasi, Fakultas Sains dan Teknologi, Universitas Jambi, Jambi, Indonesia)
Ulfa Khaira (Program Studi Sistem Informasi, Fakultas Sains dan Teknologi, Universitas Jambi, Jambi, Indonesia)
Akhiyar Waladi (Program Studi Sistem Informasi, Fakultas Sains dan Teknologi, Universitas Jambi, Jambi, Indonesia)

Jurnal Ilmu Komputer dan Sistem Komputer Terapan (JIKSTRA)
Vol. 8 No. 1 (2026): Edisi April

To address the ineffectiveness of domain blocking for illegal content such as online gambling, pornography, and piracy in Indonesia due to VPNs or new domains, this study proposes a Chromium extension based on multimodal learning. It fuses visual features using EfficientNetV2 M, text using multilingual DistilBERT, and HTML structure via early fusion into a vector of 2,187 dimensions, which is then classified by an MLP into four categories namely normal, online gambling, pornography, and piracy. Using a completely novel dataset consisting of 16,224 samples, the text and image combination achieved the best performance with an accuracy of 88.74%, a Macro F1 Score of 0.8275, and a Macro Recall of 0.812. For real world robustness, the extension utilizes all three modalities comprising text, image, and HTML, successfully classifying 36 of 40 unseen websites. Misclassifications occurred only in the digital piracy category due to limited training data and high visual similarity to legitimate websites..