yyouretoast/deepfake-detector

Dual-Stream Deepfake Detector & Spatiotemporal Forensics

A PyTorch deepfake detection architecture fusing a ConvNeXt-Small spatial backbone with a 4-stage ResSE-Spectral Tower (Steganographic Rich Model + Bayar-Stamm 2D Real FFT decomposition), High-Frequency SNR-Adaptive Gating, Platt probability calibration, and a Dual-Path (Attention + Global Max-Pooling) Bi-GRU temporal video sequence head.

• GitHub Repository: yyouretoast/deepfake-detection
• Live Hugging Face Space: yyouretoast/deepfake-detector

---

Model Description

• Developed by: Yassin
• Model Type: Dual-Stream Spatial + Frequency Hybrid Deepfake Detector
• Spatial Backbone: ConvNeXt-Small (512-d feature projection)
• Frequency Backbone: ResSE-Spectral Tower (4 stages, 20-channel SRM/Bayar 2D Real FFT input, Squeeze-and-Excitation channel attention)
• Video Temporal Head: 2-Layer Bidirectional GRU with Dual-Path Pooling (Attention + Max-Pooling)
• Calibration: Temperature Scaling ($T^* = 4.2880$) and Bayesian 3-Zone Bands ($\tau_{\text{real}}=0.40, \tau_{\text{fake}}=0.60$)

---

Hosted Checkpoints

Checkpoint	File Size	Description	SHA-256 Checksum
dual_stream_calibrated.pth	214.7 MB	Primary spatial-spectral hybrid backbone with ResSE Tower and post-hoc calibration ($T^* = 4.288$, $\tau^* = 0.4200$).	cacdd1f63a089f157d7cec384df632a37ee107830bd52a71d7faabeb50fd5237
temporal_head_best.pth	13.3 MB	2-Layer Bi-GRU video sequence head evaluating multi-frame velocity deltas and attention weights.	5976689a9132bdad51bb09cedc48c8ebbd425f353849ad15aa67073058d0700e

---

How to Use

1. Automated Hub Download & Inference

import torch
from huggingface_hub import hf_hub_download

# Download calibrated weights
ckpt_path = hf_hub_download(
    repo_id="yyouretoast/deepfake-detector",
    filename="dual_stream_calibrated.pth"
)

checkpoint = torch.load(ckpt_path, map_location="cpu", weights_only=False)
optimal_threshold = float(checkpoint.get("optimal_threshold", 0.4200))
temperature = float(checkpoint.get("temperature", 4.2880))
tau_real = float(checkpoint.get("tau_real", 0.40))
tau_fake = float(checkpoint.get("tau_fake", 0.60))

print(f"Loaded model calibrated with T*={temperature:.3f}, optimal threshold τ*={optimal_threshold:.2f}")

2. Forensic Single Image Analysis

Using the official repository engine:

from src.services.video_engine import process_single_image

# Runs YuNet 5-point face alignment, dual-stream inference, and telemetry
report = process_single_image("suspect_face.jpg")

print("Verdict:", report["three_zone"]["verdict"])
print("Calibrated Probability:", report["prob"])
print("Spectral vs Spatial Gate:", f"{report['spectral_gate']*100:.1f}% / {report['spatial_gate']*100:.1f}%")
print("SNR Attenuation (γ):", report["snr_attenuator"])

---

Empirical Benchmark Performance

Evaluated across 13,444 facial crops (756 authentic real faces, 12,688 deepfakes across 5 generator families) and 1,120 video sequences:

Metric	Single-Frame Spatial Model	Video Spatiotemporal Bi-GRU	Delta / Impact
ROC AUC	0.8248	0.8719	+4.71% discriminative improvement
PR AUC	0.8115	0.9904	Precision-recall area under curve
Equal Error Rate (EER)	24.10%	18.98%	-5.12% biometric verification error drop
Fake Precision	98.00%	98.60%	+0.60% false alarm suppression
Fake Recall	77.04%	79.75%	+2.71% detection coverage
Overall Accuracy	76.85%	79.82%	+2.97% classification rate
Expected Calibration Error (ECE)	0.0050	--	Platt-calibrated ($98.9%$ error drop from raw $0.4525$)

In-Distribution Per-Generator AUC:

• FF++ Face2Face: 0.9975
• FF++ NeuralTextures: 0.9749
• FF++ Deepfakes: 0.9625
• FF++ FaceSwap: 0.9091
• Celeb-DF v2: 0.8166

---

Citation

@misc{deepfake_forensics_2026,
  author = {Yassin},
  title = {Dual-Stream Deepfake Forensics Engine: Spatial ConvNeXt and ResSE-Spectral Gated Fusion with Spatiotemporal Sequence Modeling},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/yyouretoast/deepfake-detector}}
}