A LoRA-enhanced vision transformer for generalized and robust continual face forgery detection
In this dissertation, we present an efficient training method for face forgery detection models by combining Low-rank adaptation (LoRA) with Vision Transformers (ViT). The approach facilitates continual learning across multiple face forgery datasets organized by their release dates, followed b...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2025
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/182478 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | In this dissertation, we present an efficient training method for
face forgery detection models by combining Low-rank
adaptation (LoRA) with Vision Transformers (ViT). The
approach facilitates continual learning across multiple face
forgery datasets organized by their release dates, followed by
testing on all training datasets and an unseen dataset. Our
findings indicate that the LoRA technique significantly reduces
computation and storage costs while alleviating the issue of
catastrophic forgetting. Additionally, through experiments
varying the amount of training data, we demonstrate that the
ViT model with LoRA provides the best stability and
generalization, particularly in the context of emerging face
forgery detection techniques. |
|---|