Pengembangan Sistem Deteksi dan Klasifikasi Multimodal Disinformasi, Fitnah, dan Ujaran Kebencian pada Proyek AI Talent Factory

Kautsar, Faiz Muhammad and Ramadhan, Alif As’ad and Danutirto, Satrio Puji and Widjaya, Khairah Michiko Dyah and Wicaksono, Muhammad Farhan Arya and Gracia, Cathleen and Adhittana, Daniel and Rabbani, Nadief Aqila (2026) Pengembangan Sistem Deteksi dan Klasifikasi Multimodal Disinformasi, Fitnah, dan Ujaran Kebencian pada Proyek AI Talent Factory. Project Report. [s.n.], [s.l.]. (Unpublished)

Text 5054231013_5054231007_5054231019_5054231028_5054231011_5025231018_5025231097_5025231128-Project_Report.pdf - Accepted Version Restricted to Repository staff only Download (739kB) | Request a copy

Abstract

Penyebaran konten bermuatan Disinformasi, Fitnah, dan Ujaran Kebencian (DFK) di ruang digital Indonesia berlangsung masif, melintasi banyak platform, dan semakin kompleks karena memadukan modalitas teks, citra, audio, dan video. Kerja praktik ini dilaksanakan pada proyek AI Talent Factory DFK-3 bersama Direktorat Jenderal Pengawasan Ruang Digital (Ditjen PRD), Kementerian Komunikasi dan Digital (Komdigi), dengan tujuan membangun sistem cerdas yang mampu mengidentifikasi serta mengklasifikasikan konten DFK secara otomatis. Sistem yang dikembangkan terdiri atas dua fokus utama: model multimodal generatif berbasis arsitektur Ministral 3 yang di-fine-tune menggunakan metode LoRA untuk mengklasifikasikan konten teks dan citra ke dalam kategori Netral, Disinformasi, Fitnah, atau Ujaran Kebencian beserta trust score dan analisis naratif; serta model AI Detection berbasis framework DAViD (ViFi-CLIP) untuk membedakan konten Real, Deepfake, dan AI-Generated pada media citra, video, dan audio.

Metodologi meliputi pengumpulan dan prapemrosesan dataset hibrida berskala besar, Supervised Fine-Tuning (SFT), serta deployment model melalui endpoint API. Hasil pengujian menunjukkan model multimodal final Ministral-3-8B mencapai akurasi 94,33% dengan F1-Weighted 0,943, model DAViD mencapai akurasi 96%, dan model AI Detection Audio berbasis AST mencapai akurasi 95%, membuktikan sistem mampu menangani data multimodal dalam satu pipeline terintegrasi sejalan dengan karakteristik penyebaran konten DFK di ruang digital.
======================================================================================================================================
The spread of content containing Disinformation, Defamation, and Hate Speech (DFK — Disinformasi, Fitnah, dan Ujaran Kebencian) in Indonesia's digital space is massive in scale, spans many platforms, and is increasingly complex as it blends text, image, audio, and video modalities. This internship (Kerja Praktik) was carried out on the AI Talent Factory DFK-3 project together with the Directorate General for Digital Space Oversight (Ditjen PRD), Ministry of Communication and Digital Affairs (Komdigi), with the goal of building an intelligent system capable of automatically identifying and classifying DFK content. The developed system consists of two main components: a generative multimodal model based on the Ministral 3 architecture, fine-tuned using the LoRA method, to classify text and image content into the categories Neutral, Disinformation, Defamation, or Hate Speech, along with a trust score and narrative analysis; and an AI Detection model based on the DAViD framework (ViFi-CLIP) to distinguish between Real, Deepfake, and AI-Generated content in image, video, and audio media.

The methodology includes the collection and preprocessing of a large-scale hybrid dataset, Supervised Fine-Tuning (SFT), and model deployment via an API endpoint. Test results show that the final Ministral-3-8B multimodal model achieved 94.33% accuracy with a Weighted F1 score of 0.943, the DAViD model achieved 96% accuracy, and the AST-based Audio AI Detection model achieved 95% accuracy — demonstrating that the system is capable of handling multimodal data within a single integrated pipeline, consistent with the nature of how DFK content spreads across Indonesia's digital space.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Disinformation, Defamation, Hate Speech, Vision-Language Model, Low-Rank Adaptation , AI-Generated Content Detection, Deepfake Detection, Disinformasi, Fitnah, Ujaran Kebencian, Model Visi-Bahasa, Adaptasi Berperingkat Rendah , Deteksi Konten Buatan Kecerdasan Buatan, Deteksi Deepfake.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Satrio Puji Danutirto
Date Deposited:	12 Jul 2026 08:58
Last Modified:	12 Jul 2026 08:58
URI:	http://repository.its.ac.id/id/eprint/134722

Actions (login required)

View Item