SpectraNet: FFT-assisted Deep Learning Classifier for Deepfake Face Detection

View on arXiv ← Back to list

Authors: Nithira Jayarathne, Naveen Basnayake, Keshawa Jayasundara, Pasindu Dodampegama, Praveen Wijesinghe, Hirushika Pelagewatta, Kavishka Abeywardana, Sandushan Ranaweera, Chamira Edussooriya

Published: 2025-11-24 14:54:00+00:00

Comment: 4 pages, 3 figures

AI Summary

This paper introduces a lightweight, generalizable binary classification model named SpectraNet, based on EfficientNet-B6, for detecting deepfake face images. It addresses severe class imbalances through robust preprocessing, oversampling, and optimization strategies, achieving high accuracy and stability. While a Fourier transform-based hybrid approach was explored, its impact was minimal, with the core EfficientNet-B6 model proving most effective for accessible deepfake detection.

Abstract

Detecting deepfake images is crucial in combating misinformation. We present a lightweight, generalizable binary classification model based on EfficientNet-B6, fine-tuned with transformation techniques to address severe class imbalances. By leveraging robust preprocessing, oversampling, and optimization strategies, our model achieves high accuracy, stability, and generalization. While incorporating Fourier transform-based phase and amplitude features showed minimal impact, our proposed framework helps non-experts to effectively identify deepfake images, making significant strides toward accessible and reliable deepfake detection.

Key findings

The EfficientNet-B6 model achieved an accuracy and AUC score of 0.9102, outperforming a hybrid model that integrated Fourier transform phase and amplitude information. The integration of Fourier transform led to underperformance across all evaluated metrics, suggesting its limited utility. The EfficientNet-B6 model also demonstrated significantly lower evaluation time (2.55 seconds) compared to the hybrid model (3.48 seconds), highlighting its efficiency for real-world applications.

Approach

The authors propose using a fine-tuned EfficientNet-B6 model for binary classification of deepfake images. Their methodology includes robust data preprocessing, oversampling to address class imbalance, and advanced optimization techniques like Adam optimizer with a ReduceLROnPlateau scheduler, and mixed precision training. A hybrid model integrating Fourier Transform features with EfficientNet-B6 was also experimented with, but the core EfficientNet-B6 performed better.

Datasets

A dataset comprising 262,160 images (42,690 real, 219,470 fake) was used. Referenced datasets include Deepfakebench, Celeb-DF, FaceForensics++, and The deepfake detection challenge (dfdc) preview dataset.

Model(s)

EfficientNet-B6 for deepfake image detection. A hybrid model combining EfficientNet-B6 with Fourier Transform-based frequency analysis was also evaluated.

Author countries

Sri Lanka, Australia

← Previous