Penny-Wise and Pound-Foolish in AI-Generated Image Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• AI-generated images pose risks, necessitating effective detection methods.
• Current AI image detection models often lack long-term generalization and knowledge retention.
• Overly specialized fine-tuning on single datasets limits model effectiveness.

Purpose of the Study:

• To propose PoundNet, a novel learning framework for generalized AI-generated image detection.
• To address the trade-off between short-term gains and long-term effectiveness in detection models.
• To enhance generalization capabilities of pre-trained vision-language models for downstream detection tasks.

Main Methods:

• Developed PoundNet, a framework incorporating learnable prompt design and a balanced objective.
• Preserved broad knowledge from upstream tasks (object classification) for downstream tasks (AI-generated image detection).
• Trained PoundNet on a standard AI image dataset and evaluated on 10 large-scale public datasets.

Main Results:

• PoundNet achieved a 19% relative improvement in AI-generated image detection compared to state-of-the-art methods.
• Maintained 63% performance on object classification tasks, demonstrating balanced knowledge retention.
• Evaluated on the largest benchmark test set for AI-generated image detection generalization to date.

Conclusions:

• PoundNet effectively balances generalization and knowledge retention for AI-generated image detection.
• The proposed framework offers a more robust and effective solution for detecting AI-generated imagery.
• PoundNet represents a significant advancement in creating reliable AI image detection systems.