Reversibel wajah berbasis Dual-Decoder Autoencoder pada pengawasan ujian daring menggunakan Automatic Proctoring

Rabbani, Daffa (2026) Reversibel wajah berbasis Dual-Decoder Autoencoder pada pengawasan ujian daring menggunakan Automatic Proctoring. Jurnal Teknik Mesin, Industri, Elektro dan Informatika (JTMEI), 5 (4). pp. 14-29. ISSN 2963-7805

[img]
Preview
Text
JTMEI_VOL5_NO3_HAL_14-29_6493.pdf

Download (877kB) | Preview
Official URL: https://ejurnal.politeknikpratama.ac.id/index.php/...

Abstract

ENGLISH: Abstract. The use of online proctoring systems in Learning Management Systems continues to expand and raises privacy concerns because examinees’ faces are recorded and stored directly. This study proposes a reversible face anonymization and de-anonymization mechanism based on a Dual-Decoder Autoencoder architecture. The model consists of a single weight-sharing encoder and two separate decoders: Decoder A produces the anonymized image, while Decoder B reconstructs the original image. The model is trained using a combination of L1 (mean absolute error) and Structural Similarity Index Measure (SSIM) loss functions, weighted 100 and 1 respectively, over 1000 epochs on a real exam-proctoring face dataset split in a subject-independent manner. Evaluation shows that anonymization achieves an SSIM of 0.9874 and a Peak Signal-to-Noise Ratio (PSNR) of 42.32 decibels (dB), while de-anonymization achieves an SSIM of 0.9555 and a Peak Signal-to-Noise Ratio (PSNR) of 36.44 dB, with stable training convergence. This study emphasizes that privacy protection depends on access control over the recovery decoder rather than an internal key mechanism, so the privacy claim is stated cautiously. The proposed approach effectively conceals and restores faces with adequate visual quality to support the privacy of online exam participants. Keywords: Autoencoder; Biometric Privacy; De-Anonymization; Face Anonymization; Online Proctoring. INDONESIA: Abstrak. Penggunaan sistem pengawasan ujian daring pada Learning Management System terus meluas dan memunculkan kekhawatiran privasi karena wajah peserta direkam dan disimpan secara langsung. Penelitian ini mengusulkan mekanisme anonimisasi dan de-anonimisasi wajah yang bersifat reversibel berbasis arsitektur DualDecoder Autoencoder. Model terdiri atas satu encoder dengan bobot bersama dan dua decoder terpisah: Decoder A menghasilkan citra teranonimisasi, sedangkan Decoder B merekonstruksi citra asli. Model dilatih menggunakan kombinasi fungsi loss L1 (mean absolute error) dan Structural Similarity Index Measure (SSIM) dengan bobot masing-masing 100 dan 1 selama 1000 epoch pada dataset wajah pengawasan ujian nyata yang dibagi secara subject-independent. Hasil evaluasi menunjukkan bahwa anonimisasi mencapai SSIM 0,9874 dan Peak Signalto-Noise Ratio (PSNR) 42,32 desibel (dB), sedangkan de-anonimisasi mencapai SSIM 0,9555 dan PSNR 36,44 dB dengan konvergensi pelatihan yang stabil. Penelitian ini menegaskan bahwa perlindungan privasi bergantung pada kontrol akses terhadap decoder pemulih, bukan pada mekanisme kunci internal, sehingga klaim privasi dirumuskan secara hati-hati. Pendekatan yang diusulkan mampu menyamarkan dan memulihkan wajah dengan kualitas visual yang memadai untuk mendukung privasi peserta ujian daring.

Item Type: Article
Uncontrolled Keywords: Anonimisasi Wajah; Autoencoder; De-Anonimisasi; Pengawasan Daring; Privasi Biometrik;
Subjects: Data Processing, Computer Science > Computer and Human
Special Computer Methods > Artificial Intelligence
Divisions: Fakultas Sains dan Teknologi > Program Studi Teknik Informatika
Depositing User: Daffa Rabbani
Date Deposited: 04 Aug 2026 04:19
Last Modified: 04 Aug 2026 04:19
URI: https://digilib.uinsgd.ac.id/id/eprint/137504

Actions (login required)

View Item View Item