Implementasi Federated Learning Pada Sistem Kehadiran Berbasis Pengenalan Wajah Di Lingkungan Edge Computing

Figo, Steven (2026) Implementasi Federated Learning Pada Sistem Kehadiran Berbasis Pengenalan Wajah Di Lingkungan Edge Computing. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5027221021-Undergraduate_Thesis.pdf Download (8MB)

Abstract

Sistem absensi berbasis pengenalan wajah saat ini telah banyak diterapkan untuk mengotomatiskan pencatatan kehadiran, namun metode konvensional berbasis Centralized Learning mengharuskan pengumpulan citra wajah mentah ke server pusat yang rentan terhadap risiko kebocoran data biometrik sensitif serta pemborosan bandwidth jaringan. Penelitian ini mengusulkan implementasi Federated Learning menggunakan algoritma FedProx dan model MobileFaceNet yang dikontainerisasi dengan Docker untuk membangun sistem kehadiran terdistribusi pada lingkungan Edge Computing menggunakan perangkat keras heterogen (Jetson Nano dan Raspberry Pi 3 Model B). Pendekatan ini memungkinkan pelatihan model dilakukan secara lokal pada masing-masing perangkat edge tanpa mengirimkan data biometrik mentah mahasiswa ke server, sehingga privasi data tetap terjaga. Pengujian dilakukan melalui evaluasi performa pelatihan model, efisiensi sumber daya, serta pengujian inferensi nyata dan video simulasi. Hasil penelitian menunjukkan bahwa model global hasil Federated Learning mencapai akurasi validasi hingga 99,69%, sebanding bahkan sedikit melampaui Centralized Learning yang konsisten mencatatkan akurasi validasi 99,65%. Dari sisi efisiensi sumber daya, Federated Learning jauh lebih hemat bandwidth dengan hanya mentransmisikan 167,28 MB, sekitar 22% dari kebutuhan Centralized Learning sebesar 744,99 MB, namun harus dibayar dengan durasi pelatihan dan konsumsi energi lebih besar akibat keterbatasan perangkat Raspberry Pi. Pada pengujian inferensi nyata, sistem mencatatkan keberhasilan fungsionalitas 100% pada jarak 0,5 hingga 3 meter di kondisi cahaya cukup maupun minim, dengan akurasi klasifikasi stabil di atas 90% menggunakan ambang batas kepercayaan 0,7. Latensi inferensi lokal pada Jetson Nano tercatat cepat sekitar 700-765 ms, sedangkan Raspberry Pi jauh lebih lambat sekitar 3100-3230 ms akibat keterbatasan RAM. Meskipun demikian, pengujian video simulasi menunjukkan Federated Learning masih lebih rentan terhadap kasus salah absen serta dibandingkan Centralized Learning. Penelitian ini menunjukkan bahwa integrasi Federated Learning dan Edge Computing efektif membangun sistem presensi yang hemat bandwidth dan menjaga privasi data biometrik, meski masih memerlukan peningkatan ketahanan terhadap kesalahan identifikasi.
==============================================================================================================================
Face recognition-based attendance systems have been widely adopted to automate presence recording; however, conventional Centralized Learning methods require collecting raw facial images at a central server, posing risks of biometric data leakage and excessive network bandwidth consumption. This study proposes implementing Federated Learning using the FedProx algorithm and MobileFaceNet model, containerized with Docker, to build a distributed attendance system in an Edge Computing environment using heterogeneous hardware (Jetson Nano and Raspberry Pi 3 Model B). This approach enables local model training on each edge device without transmitting raw student biometric data to the server, thereby preserving data privacy. Evaluation was conducted through training performance assessment, resource efficiency analysis, and real-life inference and video simulation testing. Results show the global model produced by Federated Learning achieved a validation accuracy of up to 99.69%, comparable to and slightly exceeding the consistent 99.65% accuracy of Centralized Learning. Federated Learning also demonstrated significantly greater bandwidth efficiency, transmitting only 167.28 MB, around 22% of the 744.99 MB required by Centralized Learning, though at the cost of longer training duration and higher energy consumption due to the limited capacity of the Raspberry Pi. In real-life inference testing, the system achieved 100% functional success at distances of 0.5 to 3 meters under both adequate and low-light conditions, with classification accuracy consistently above 90% at a confidence threshold of 0.7. Local inference latency on the Jetson Nano remained fast at approximately 700-765 ms, while the Raspberry Pi was considerably slower at approximately 3,100-3,230 ms due to RAM constraints. Nevertheless, video simulation testing revealed that Federated Learning remains more susceptible to false acceptance cases than Centralized Learning. These findings demonstrate that Federated Learning and Edge Computing effectively build a bandwidth-efficient, privacy-preserving attendance system, though further improvement in robustness against misidentification is still required.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Edge Computing, Face Recognition, Federated Learning, MobileFaceNet, Smart Attendance.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > Q Science (General) > Q337.5 Pattern recognition systems Q Science > QA Mathematics > QA336 Artificial Intelligence T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.B56 Biometric identification
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Steven Figo
Date Deposited:	14 Jul 2026 07:06
Last Modified:	14 Jul 2026 07:06
URI:	http://repository.its.ac.id/id/eprint/134894

Actions (login required)

View Item