![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-3-4-e1712870914954.png)
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-4-e1712870974203.png)
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/veritaslightlogoblue-300x105.png)
User Analysis
Analysis of our primary users: AI researchers and journalists.
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-5-1.png)
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-6-1.png)
Main features chosen after user analysis: Retrainability, Ease of use, Explainability and Relevance
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-18-1024x576.png)
Inference
Verify the authenticity of audio files by uploading it and running an inference process.
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/Upload_file-1-1.png)
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/Inferance-3-1.png)
Select models based on detailed information provided on Veritas.
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-19-1-1024x648.png)
Retraining
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/image-6-1024x576.png)
Retraining models for new deepfake audio types is now simpler. Upload a zip file with the dataset, adjust parameters, and start training
VeritasAlpha achieves state-of-the-art performance, merging strengths of different models, enhancing EER by 30%.
VeritasSG is trained for the Singaporean accent using our retraining feature. It detects Singaporean accent deepfakes better than current market models.
VeritasAlpha achieved state-of-the-art performance with the lowest Equal Error Rate (EER) seen. VeritasSG achieves an EER of 0 with the Singaporean Accent audio file.
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/Equal_Error_rate-1-1-1536x864.png)
Software Architecture
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/8-7.png)
![](https://capstoneshowcase.sutd.edu.sg/wp-content/uploads/2024/03/7-7.png)
User Feedback
We collected responses from approximately 100 users, comprising 50% AI researchers, 40% general public, and 10% journalists using Maze
Found it easy to authenticate audio file
Found highlighting potential deepfaked portions useful
Navigated through the software to retrain models easily
-
Researcher 1
"The detection and model retraining feature offered significantly contributes to streamlining my work processes."
-
Researcher 2
"Interface for initiating retraining is intuitive and straightforward, making it accessible even for users with minimal technical expertise."
-
Journalist
"Overall, I find this feature to be highly beneficial and user-friendly, effectively empowering me to achieve better outcomes in my work."
-
General Public
"I think it works well for detecting deepfakes especially when critical information is being shared. Would be useful with financial information."