ML Security PapersSpectraNet: FFT-assisted Deep Learning Classifier for Deepfake Face Detection
Nithira Jayarathne 1, Naveen Basnayake 1, Keshawa Jayasundara 1, Pasindu Dodampegama 1, Praveen Wijesinghe 1, Hirushika Pelagewatta 1, Kavishka Abeywardana 1, Sandushan Ranaweera 2, Chamira Edussooriya 1
Published on arXiv
2511.19187
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
EfficientNet-B6 achieves 91% accuracy on deepfake detection; FFT-based phase/amplitude features provide negligible additional benefit over the CNN baseline alone.
SpectraNet
Novel technique introduced
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 Contributions
- Lightweight EfficientNet-B6 binary classifier fine-tuned for deepfake face detection with balanced epoch-level batch sampling to handle severe class imbalance (42K real vs 219K fake images)
- Hybrid SpectraNet framework integrating Fourier transform phase/amplitude features alongside CNN features, with empirical finding that FFT contributes minimally to detection accuracy
- Comprehensive evaluation across multiple deepfake datasets (FaceForensics++, Celeb-DF, DFDC) demonstrating 91% accuracy with strong generalization
🛡️ Threat Analysis
Directly addresses AI-generated content detection by proposing a novel binary classifier to distinguish real from deepfake face images — a canonical output integrity / content authenticity task.