Unlocking the Potential of Lightweight Quantized Models for Deepfake Detection

Renshuai Tao, Ziheng Qin, Yifu Ding, Chuangchuang Tan, Jiakai Wang, Wei WangIJCAI 2025 · DOI

Introduction

As deepfake technology becomes increasingly sophisticated, the need for real-time detection on edge devices—such as mobile phones and IoT sensors—has grown critical. However, current state-of-the-art detection models rely on heavyweight architectures like Transformers or deep ResNets, which are too computationally expensive and memory-intensive for resource-constrained environments.

While quantization (reducing precision from 32-bit floats to low-bit integers) is a standard solution for model compression, it often fails in deepfake detection. This task relies on subtle, high-frequency texture artifacts that are easily lost during the rounding process of standard quantization. This paper introduces QMDD (Quantized Model for Deepfake Detection), a framework designed to maintain high accuracy even at extremely low bit-widths.

Methodology

The QMDD framework overcomes the limitations of standard quantization through three primary innovations:

1. NPR (Neighboring Pixel Relationships) Representation

Instead of processing raw RGB images, the model utilizes NPR. This representation emphasizes upsampling artifacts—common signatures of AI generation—while remaining robust to the noise introduced by quantization. As shown in comparative evaluations, NPR features maintain significantly higher accuracy than raw image features as the bit-width decreases.

Accuracy across different bit-widths comparing raw image features vs. NPR features.

2. Connected Quantized Block (CQB)

To prevent the loss of fine-grained details, the authors developed the CQB. This modified residual block employs a quantized main path for general feature extraction alongside a full-precision auxiliary shortcut. This shortcut uses computationally inexpensive operations like pooling to preserve method-specific textures that low-bit quantization might otherwise destroy.

3. Shifted Logarithmic Redistribution (SLR) Quantizer

Standard uniform quantization is inefficient for neural network activations, which are typically clustered near zero. The SLR quantizer applies a logarithmic transformation to "unfold" these activations before quantization. This allows for finer granularity in high-density regions without adding computational overhead, as it can be implemented via a simple lookup table during inference.

Evaluation and Results

The framework was tested on major benchmarks including ForenSynths (Cross-GAN evaluation) and UniversalFakeDetect (Cross-Diffusion evaluation). The results demonstrate that massive models are often redundant for the binary task of deepfake detection.

Efficiency and Accuracy Gains

QMDD achieves a 10.8× reduction in computational costs (GFLOPs) and a 12.4× reduction in storage compared to a standard ResNet-50. Remarkably, a 2-bit quantized version of the model achieved 91.8% accuracy on the ForenSynths benchmark, outperforming the state-of-the-art CLIP-based detector (89.1%) while using less than 5% of its computational resources.

Comparison of accuracy, storage (MB), and computational cost (GFLOPs) between CLIP, ResNet-50, and the proposed model.

Generalization

The model maintains strong performance on "unseen" domains—generators not present in the training set—proving that the learned quantized features are robust indicators of forgery rather than artifacts of specific training noise.

Conclusion

By making quantization "aware" of forgery-specific features through NPR and SLR, the QMDD framework proves that lightweight models can match or exceed the performance of heavyweight counterparts. While future work is needed to address temporal consistency in video deepfakes and evolving generative techniques, this research provides a viable path for deploying high-security deepfake detection on the edge.

Summary updated

BibTeX
@inproceedings{tao2025unlocking,
  author = {Renshuai Tao and Ziheng Qin and Yifu Ding and Chuangchuang Tan and Jiakai Wang and Wei Wang},
  title = {{Unlocking the Potential of Lightweight Quantized Models for Deepfake Detection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/59},
}
Figure