Penny-Wise and Pound-Foolish in Deepfake Detection

View on arXiv ← Back to list

Authors: Yabin Wang, Zhiwu Huang, Su Zhou, Adam Prugel-Bennett, Xiaopeng Hong

Published: 2024-08-15 20:38:31+00:00

AI Summary

This paper addresses the problem of deepfake detection model generalization by proposing PoundNet, a novel learning framework that incorporates a learnable prompt design and a balanced objective function. PoundNet significantly improves deepfake detection performance (19% improvement) while maintaining strong object classification performance (63%), unlike other state-of-the-art methods.

Abstract

The diffusion of deepfake technologies has sparked serious concerns about its potential misuse across various domains, prompting the urgent need for robust detection methods. Despite advancement, many current approaches prioritize short-term gains at expense of long-term effectiveness. This paper critiques the overly specialized approach of fine-tuning pre-trained models solely with a penny-wise objective on a single deepfake dataset, while disregarding the pound-wise balance for generalization and knowledge retention. To address this Penny-Wise and Pound-Foolish issue, we propose a novel learning framework (PoundNet) for generalization of deepfake detection on a pre-trained vision-language model. PoundNet incorporates a learnable prompt design and a balanced objective to preserve broad knowledge from upstream tasks (object classification) while enhancing generalization for downstream tasks (deepfake detection). We train PoundNet on a standard single deepfake dataset, following common practice in the literature. We then evaluate its performance across 10 public large-scale deepfake datasets with 5 main evaluation metrics-forming the largest benchmark test set for assessing the generalization ability of deepfake detection models, to our knowledge. The comprehensive benchmark evaluation demonstrates the proposed PoundNet is significantly less Penny-Wise and Pound-Foolish, achieving a remarkable improvement of 19% in deepfake detection performance compared to state-of-the-art methods, while maintaining a strong performance of 63% on object classification tasks, where other deepfake detection models tend to be ineffective. Code and data are open-sourced at https://github.com/iamwangyabin/PoundNet.

Key findings

PoundNet achieves a 19% average improvement in deepfake detection performance across 10 datasets compared to state-of-the-art methods. It also maintains a strong 63% accuracy on object classification tasks, unlike other deepfake detection models which perform poorly on these tasks. The results demonstrate that PoundNet's balanced approach is significantly less prone to overfitting and catastrophic forgetting.

Approach

PoundNet fine-tunes a pre-trained vision-language model (CLIP) using a learnable prompt design and a balanced objective function. This objective combines a class-agnostic binary loss, a semantic-preserving term for knowledge retention, and a class-aware binary term to improve generalization.

Datasets

ForenSynths (for training), 10 public large-scale deepfake datasets (for testing; including ForenSynths, GenImage, GANGen-Detection, DiffusionForensics, Ojha, AntiFake, DIF, UADFV, Celeb-DF v1 and v2), and 5 ImageNet datasets (ImageNet, ImageNet-A, ImageNet-V2, ImageNet-R, ImageNet-S) for object classification.

Model(s)

CLIP (Contrastive Language-Image Pre-training) model, specifically ViT-L/14.

Author countries

China, United Kingdom

← Previous