Perlindungan Anak di Ruang Digital (PAD) Computer Vision Program AI Talent Factory (AITF)

Nurrahman, Muhammad Daffa and Martin, Filbert Hainsly and Arifin, Muhammad Naufal and Wicaksono, Dimas Ahmad Satrio and Kurniawan, Dhafin (2026) Perlindungan Anak di Ruang Digital (PAD) Computer Vision Program AI Talent Factory (AITF). Project Report. [s.n.], [s.l]. (Unpublished)

Text 5025231208_5025231256_5054231006_5054231015_5054231016-Project_Report.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Perkembangan ruang digital telah menghadirkan risiko yang semakin kompleks bagi anak, mulai dari paparan konten pornografi dan kekerasan, cyberbullying, hingga praktik grooming yang sulit dideteksi oleh sistem moderasi konvensional berbasis kata kunci. Sebagai respons atas Peraturan Pemerintah Nomor 17 Tahun 2025 tentang Tata Kelola Penyelenggaraan Sistem Elektronik dalam Perlindungan Anak (PAD), Tim 3 Artificial Intelligence Talent Factory (AITF) Institut Teknologi Sepuluh Nopember mengembangkan sistem Perlindungan Anak di Ruang Digital (PAD), yaitu platform kecerdasan buatan multimodal untuk memantau, mengklasifikasikan, dan memberikan reasoning atas konten digital berupa teks, gambar, dan video yang berpotensi membahayakan anak. Sistem mengintegrasikan Large Language Model (LLM), Vision Language Model (VLM), Computer Vision, Agentic AI, dan Retrieval-Augmented Generation (RAG) dalam arsitektur berlapis, mencakup crawling dan keyword generation, klasifikasi multimodal, reasoning berbasis regulasi. Dua skenario pelatihan model diuji dan dibandingkan: model terpisah yang digabungkan melalui adapter merging, serta model tunggal (unified VLM) yang dilatih secara multi-task. Hasil pengujian menunjukkan bahwa pendekatan adapter merging menyebabkan penurunan performa yang drastis(catastrophic degradation), sementara pendekatan model tunggal berhasil mempertahankan akurasi tinggi secara konsisten pada seluruh tugas klasifikasi teks (92,76%), gambar (91,21%), dan keyword generator (92,61%). Evaluasi end-to-end terhadap logika Agentic AI, kualitas RAG, dan fungsionalitas dashboard admin turut dilakukan untuk memvalidasi kesiapan sistem pada lingkup 14 Minimum Viable Product (MVP). Hasil kerja praktik ini menunjukkan bahwa pendekatan kecerdasan buatan multimodal berbasis model tunggal lebih andal dibandingkan penggabungan model terpisah dalam mendukung tata kelola ruang digital yang aman dan ramah anak sesuai ketentuan PAD.
==================================================================================================================================
The rapid development of the digital environment has introduced increasingly complex risks for children, including exposure to pornographic and violent content, cyberbullying, and online grooming practices that are difficult to detect using conventional keyword-based moderation systems. In response to Government Regulation of the Republic of Indonesia Number 17 of 2025 concerning the Governance of Electronic System Providers for Child Protection (Perlindungan Anak Digital/PAD), Team 3 of the Artificial Intelligence Talent Factory (AITF) at Institut Teknologi Sepuluh Nopember developed the Digital Child Protection (PAD) system, an artificial intelligence-based multimodal platform designed to monitor, classify, and provide reasoning for digital content in the form of text, images, and videos that may pose risks to children. The system integrates Large Language Models (LLMs), Vision Language Models (VLMs), Computer Vision, Agentic AI, and Retrieval-Augmented Generation (RAG) within a layered architecture consisting of web crawling and keyword generation, multimodal content classification, and regulation-based reasoning. Two model training strategies were implemented and evaluated: separate task-specific models combined through adapter merging and a unified Vision Language Model trained using a multi-task learning approach. Experimental results indicate that the adapter merging strategy suffered from catastrophic degradation, leading to a significant decline in overall performance. In contrast, the unified model consistently achieved high classification accuracy across all tasks, including text classification (92.76%), image classification (91.21%), and keyword generation (92.61%). End-to-end evaluations were also conducted to assess the Agentic AI decision-making logic, the quality of the Retrieval-Augmented Generation (RAG) module, and the functionality of the administrative dashboard, validating the system's readiness across all 14 defined Minimum Viable Product (MVP) features. The results demonstrate that a unified multimodal artificial intelligence model provides a more robust and reliable solution than combining separate models through adapter merging, thereby offering effective support for creating a safer and more child-friendly digital environment in accordance with the PAD regulatory framework.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Perlindungan Anak di Ruang Digital, PAD, Kecerdasan Buatan Multimodal, Vision Language Model, Agentic AI
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Daffa Nurrahman
Date Deposited:	06 Jul 2026 06:56
Last Modified:	06 Jul 2026 06:56
URI:	http://repository.its.ac.id/id/eprint/134346

Actions (login required)

View Item