A New Deepfake Detection Method with No-Reference Image Quality Assessment to Resist Image Degradation

Abstract

Deepfake technology, which utilizes advanced AI models such as Generative Adversarial Networks (GANs), has led to the proliferation of highly convincing manipulated media, posing significant challenges for detection. Existing detection methods often struggle with the low-quality or compressed press, which is prevalent on social media platforms. This paper proposes a novel Deepfake detection framework that leverages No-Reference Image Quality Assessment (NRIQA) techniques, specifically, BRISQUE, NIQE, and PIQUE, to extract quality-related features from facial images. These features are then classified using a Support Vector Machine (SVM) with various kernel functions. We evaluate our method under both intra-dataset and cross-dataset settings. For intra-dataset evaluation, we conduct K-fold cross-validation on two benchmark datasets, DFDC and Celeb-DF (v2), including downsampled versions to simulate real-world degradation. The results show that our method maintains high accuracy even under significant quality loss, achieving up to 98% accuracy on the Celeb-DF (v2) dataset and outperforming several state-of-the-art methods. To improve the transferability of the detection models, we introduce an integrated filtering strategy based on NR-IQA thresholding, which enhances performance in cross-dataset transfer scenarios. This approach yields up to 7% improvement in detection accuracy under challenging cross-domain conditions.