Diash, Hakam Dzakwan (2026) Klasifikasi Audio Deepfake dengan Ekstraksi Fitur YAMNet dan Deep Neural Network. Undergraduate thesis, UPN Veteran Jawa Timur.
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (Cover)
22083010069-COVER.pdf Download (1MB) |
|
Text (Bab 1)
22083010069-BAB I.pdf Download (231kB) |
|
Text (Bab 2)
22083010069-BAB II.pdf Restricted to Repository staff only until 13 July 2029. Download (594kB) |
|
Text (Bab 3)
22083010069-BAB III.pdf Restricted to Repository staff only until 13 July 2029. Download (677kB) |
|
Text (Bab 4)
22083010069-BAB IV.pdf Restricted to Repository staff only until 13 July 2029. Download (1MB) |
|
Text (Bab 5)
22083010069-BAB V.pdf Download (222kB) |
|
Text (Daftar Pustaka)
22083010069-DAFTAR PUSTAKA.pdf Download (180kB) |
|
Text (Lampiran)
22083010069-LAMPIRAN.pdf Restricted to Repository staff only Download (392kB) |
Abstract
The advancement of artificial intelligence technology in voice synthesis has increased the capability of generating deepfake audio that is difficult to distinguish from genuine human speech, potentially leading to misinformation, identity fraud, and misuse of digital media. This study aims to evaluate the effectiveness of combining YAMNet feature extraction and Deep Neural Network (DNN) for deepfake audio classification, analyze the application of statistical aggregation methods on YAMNet embeddings, measure cross-dataset generalization capability, and implement the trained model into a Flask-based web prototype. The research method was conducted by extracting audio embeddings using YAMNet, followed by applying statistical aggregation methods such as mean and standard deviation to generate more compact feature representations before being used as input for the DNN classification model. The results show that statistical aggregation methods successfully reduced data complexity while maintaining high classification performance with accuracy ranging from 97.58% to 99.24%, where the best model was achieved by the Mean + Std Model 1 method with an accuracy of 99.24%. In cross-dataset testing, the initial model demonstrated low performance with primary dataset accuracy ranging from 61.00% to 62.37%, however, after fine-tuning using the freeze layer approach, the model achieved a more balanced performance with primary dataset accuracy of 92.81% and secondary dataset accuracy of 91.41%. Furthermore, the trained model was successfully implemented into a Flask-based system called VocalSentry, which provides interactive deepfake audio detection features through an informative and user-friendly interface. Based on the results, the combination of YAMNet feature extraction and DNN proved effective for deepfake audio classification and was able to maintain good performance after cross-dataset fine-tuning.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > QA Mathematics > QA76.6 Computer Programming | ||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Data Science | ||||||||||||
| Depositing User: | Hakam Dzakwan Diash | ||||||||||||
| Date Deposited: | 14 Jul 2026 06:30 | ||||||||||||
| Last Modified: | 14 Jul 2026 06:30 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/55296 |
Actions (login required)
 |
View Item |