Detection Models

DFPN ships with four pre-configured detection models that cover the most common deepfake and synthetic-media threats. Each model is maintained as a standalone module under the models/ directory and can be run by any worker node.

---

Pre-configured Models

face-forensics -- Face Manipulation Detection

Detects face-swap and reenactment forgeries in still images using the Self-Blended Images (SBI) training strategy on an EfficientNet-B4 backbone.

Benchmark	Accuracy
FaceForensics++ (c23)	97.2%
Celeb-DF v2	91.4%

Processing speed: ~50 ms on GPU / ~500 ms on CPU

Best for

Profile photos, identity documents, social-media headshots -- any image where a face is the primary subject.

---

universal-fake-detect -- AI-Generated Image Detection

A CLIP-based classifier (CLIP-ViT-L/14) fine-tuned to separate real photographs from outputs of modern generative models.

Generator	Accuracy
ProGAN	99.8%
Stable Diffusion	89.4%
DALL-E	85.7%

Processing speed: ~100 ms on GPU / ~800 ms on CPU

Best for

Identifying fully synthetic images produced by diffusion models, GANs, or other generative pipelines.

---

video-ftcn -- Video Authenticity Detection

Combines an Xception frame-level feature extractor with a Temporal CNN to capture inter-frame inconsistencies that reveal video-level manipulation.

Benchmark	Accuracy
FaceForensics++	96.4%
Celeb-DF v2	88.9%

Processing speed: ~2 s on GPU / ~30 s on CPU

CPU note

Video analysis is extremely slow on CPU-only nodes. The default CPU configuration disables this modality. See CPU-only configuration for details.

---

ssl-antispoofing -- Voice Cloning Detection

Leverages wav2vec 2.0 / XLSR-53 self-supervised speech representations to detect synthetic and cloned voices.

Benchmark	Accuracy
ASVspoof 2021 (LA)	99.2%

Processing speed: ~200 ms on GPU / ~2 s on CPU

Best for

Voice messages, phone-call recordings, podcast clips -- any audio where speaker authenticity matters.

---

Performance Comparison

Accuracy by model

Model	Primary Benchmark	Accuracy	Secondary Benchmark	Accuracy
face-forensics	FF++ (c23)	97.2%	Celeb-DF v2	91.4%
universal-fake-detect	ProGAN	99.8%	Stable Diffusion	89.4%
video-ftcn	FF++	96.4%	Celeb-DF v2	88.9%
ssl-antispoofing	ASVspoof 2021	99.2%	--	--

Processing speed

Model	GPU Latency	CPU Latency	GPU Required?
face-forensics	50 ms	500 ms	Recommended
universal-fake-detect	100 ms	800 ms	Recommended
video-ftcn	2 s	30 s	Strongly recommended
ssl-antispoofing	200 ms	2 s	Recommended

---

Supported Modalities

Every model, worker, and analysis request declares which modalities it supports using a bitfield. This allows efficient on-chain filtering without string comparison.

Modality	Description	Bit Value
ImageAuthenticity	Real vs. AI-generated image classification	1 (bit 0)
VideoAuthenticity	Temporal forgery detection in video	2 (bit 1)
AudioAuthenticity	General audio manipulation detection	4 (bit 2)
FaceManipulation	Face-swap and reenactment detection	8 (bit 3)
VoiceCloning	Synthetic / cloned voice detection	16 (bit 4)
GeneratedContent	Fully AI-generated media detection	32 (bit 5)

Combine values with bitwise OR. For example, a worker that handles face manipulation and AI-generated images advertises modalities = 8 | 1 | 32 = 41.

Model-to-modality mapping

Model	Modalities
face-forensics	FaceManipulation (8)
universal-fake-detect	ImageAuthenticity (1) + GeneratedContent (32) = 33
video-ftcn	VideoAuthenticity (2)
ssl-antispoofing	VoiceCloning (16)

---

Standardized Output Format

All models produce a JSON result with the same envelope so that workers and on-chain aggregation logic can treat them uniformly.

{
  "verdict": "manipulated",
  "confidence": 0.973,
  "detections": [
    {
      "type": "face_swap",
      "confidence": 0.973,
      "region": {
        "x": 120,
        "y": 80,
        "width": 256,
        "height": 256
      },
      "metadata": {
        "model": "face-forensics-sbi",
        "version": "1.0.0"
      }
    }
  ]
}

Field	Type	Description
verdict	string	One of authentic, manipulated, or inconclusive
confidence	float	Overall confidence score between 0.0 and 1.0
detections	array	Individual findings, each with its own confidence and optional spatial region
detections[].type	string	Detection category (e.g. face_swap, generated_image, voice_clone)
detections[].region	object	Bounding box for spatial detections (images/video); omitted for audio
detections[].metadata	object	Model identifier and version that produced this detection