top of page
Annaverse-logo

Deepfake Detection

Overview

Paper PDF File Preview: Click me

Timeline: Sep 2024 – Jun 2025
Role: AI Researcher
Team: 2 members, 1 instructor
Tools: Python, AI models
Type: Research Project (Conference Presentation)

This project investigates how well AI models detect deepfake images across different cultural datasets.

Deepfake Detection Overview Image
Problem

Most deepfake detection models:

  • Perform well on familiar datasets
  • Fail when applied to different populations

This raises concerns about:

  • Bias
  • Reliability
  • Fairness
Goal

Evaluate how well models generalize across:

  • East Asian faces (TWHD dataset)
  • Western faces (FaceForensics++)
Methodology
語言模型概念

Models Used:

  • Xception

  • MobileNetV3

  • Vision Transformer (ViT)

Approach:

  • Train and test models on different datasets

  • Compare self-validation vs cross-dataset performance

Key Findings

  • High accuracy within same dataset

  • Significant performance drop in cross-cultural testing

Root Causes

  • Demographic bias

  • Background/environment differences

Outcome

  • Highlighted limitations of current AI models

  • Emphasized need for more diverse datasets

Reflection

Key Learning

  • AI performance is highly context-dependent
  • Bias can significantly impact real-world usability

Product Thinking

  • AI systems must be designed with fairness and inclusivity
  • Data diversity is as important as model performance
PREVIOUS PROJECT
BACK TO WORK
NEXT PROJECT
bottom of page