SPECTRAFORGE: Domain-Equalized Frequency-Spatial Fusion for Synthetic Dermatology Detection
Abstract
Spatial deepfake detectors in medical imaging often exploit dataset-level artifacts rather than authentic generative fingerprints, leading to severe performance degradation on unseen generators. We introduce SPECTRAFORGE, a two-stream CNN framework that decouples spatial morphology from frequency-domain synthetic artifacts. A Gaussian-bottlenecked spatial stream captures lesion morphology, while a parallel FFT-magnitude stream targets periodic upsampling artifacts from diffusion decoders. An Extreme Equalizer preprocessing pipeline is used to eliminate spatial dataset leakage before feature extraction. On a controlled 2,000-image forensic cohort, SPECTRAFORGE achieves an AUC of 0.9971 ± 0.0016 and Precision of 0.9931 ± 0.0056. Under cross-checkpoint OOD evaluation, SPECTRAFORGE maintains an AUC of 0.9277 compared with 0.5494 for EfficientNet-B0, demonstrating substantially stronger domain generalization.
Key Methodologies & Contributions
- Extreme Equalizer Preprocessing: Developed an in-memory preprocessing pipeline using grayscale conversion, border cropping, and Gaussian filtering to reduce dataset-level spatial leakage and artifact biases.
- Dual-Stream CNN Architecture: Designed parallel spatial and frequency streams using a Gaussian-bottlenecked morphology pathway and an FFT-magnitude pathway for synthetic artifact detection.
- Robust OOD Generalization: Achieved an OOD AUC of 0.9277 under cross-checkpoint evaluation, substantially outperforming the EfficientNet-B0 baseline.
- High-Precision Forensic Detection: Achieved 0.9971 ± 0.0016 AUC and 0.9931 ± 0.0056 Precision across three-seed cross-validation on a controlled 2,000-image cohort.
Publication Status
Accepted for Short Oral Presentation at IEEE DSAA, 2026 — CORE A Tier (<25% acceptance rate)
Authors: A. Kumar, L. Chhetri, D. Das
- All authors contributed equally in this research