VeridisQuo

Where is the truth? What is Veridis Quo? A state-of-the-art neural network designed to detect deepfake videos and highlight the altered areas in each frame using explainable AI. Try it online!
Result

Pipeline

%%{init: {'theme': 'base', 'themeVariables': { 'primaryColor': '#1a1a2e', 'primaryTextColor': '#fff', 'lineColor': '#ffffff', 'background': '#0d1117', 'mainBkg': '#0d1117'}}}%%

flowchart TB
    subgraph Preprocess[" Preprocess "]
        A[MP4 Video] --> B[Frames extraction]
        B --> C[Faces detection]
        C --> D[Faces extraction]
    end

    D --> E
    D --> G
    D --> H

    subgraph Processing[" "]
        direction LR
        subgraph Spatial_Module[" Spatial Module "]
            E[Features ExtractionEfficientNet] -->|"1792, 7, 7"| F((Pooled))
        end

        subgraph Frequency_Module[" Frequency Module "]
            G[DCT Extractor8×8, freq_bands] -->|"512"| I((Concat))
            H[FFT Extractorradial=8, hann] -->|"512"| I
            I -->|"1024"| J[FusionMLP1024→512→1024]
        end
    end

    F -->|"1792"| K
    J -->|"1024"| K

    subgraph Classifier_Module[" Classifier Module "]
        K((Concat)) -->|"2816"| L[MLP]
        L --> M[Frames Aggregation]
        M --> N[Final Score]
    end

    E --> O
    D --> Q

    subgraph Gradcam_Module[" Gradcam Module "]
        O[Gradcam Computation] --> P[Gradcam Visualization]
        P --> Q((Remap))
        Q --> R[Final Manipulated Video]
    end

    %% Styling
    classDef preprocess fill:#1a1a2e,stroke:#f59e0b,color:#f59e0b
    classDef spatial fill:#1a1a2e,stroke:#10b981,color:#10b981
    classDef frequency fill:#1a1a2e,stroke:#3b82f6,color:#3b82f6
    classDef classifier fill:#1a1a2e,stroke:#ec4899,color:#ec4899
    classDef gradcam fill:#1a1a2e,stroke:#a855f7,color:#a855f7
    classDef default fill:#1a1a2e,stroke:#6b7280,color:#fff

    class B,C,D preprocess
    class E,F spatial
    class G,H,I,J frequency
    class K,L,M,N classifier
    class O,P,Q,R gradcam

    style Preprocess fill:#0d1117,stroke:#f59e0b,color:#f59e0b
    style Spatial_Module fill:#0d1117,stroke:#10b981,color:#10b981
    style Frequency_Module fill:#0d1117,stroke:#3b82f6,color:#3b82f6
    style Classifier_Module fill:#0d1117,stroke:#ec4899,color:#ec4899
    style Gradcam_Module fill:#0d1117,stroke:#a855f7,color:#a855f7
    style Processing fill:transparent,stroke:none

Loading

---

Model Architecture

Hybrid Detection System

graph LR
    Input["Input Image<br/>224×224×3"] --> Spatial["<b>Spatial Stream</b><br/>EfficientNet-B4<br/>(ImageNet)"]
    Input --> Frequency["<b>Frequency Stream</b>"]

    Frequency --> FFT["FFT Extractor<br/>8 radial bands<br/>Hann window"]
    Frequency --> DCT["DCT Extractor<br/>8×8 blocks<br/>frequency bands"]

    FFT --> FFT_Out["512-dim"]
    DCT --> DCT_Out["512-dim"]

    FFT_Out --> Fusion["Fusion MLP"]
    DCT_Out --> Fusion

    Spatial --> Spatial_Out["1792-dim"]
    Fusion --> Fusion_Out["1024-dim"]

    Spatial_Out --> Concat["Concatenate<br/>2816-dim"]
    Fusion_Out --> Concat

    Concat --> Classifier["<b>Classifier MLP</b><br/>1024 → 512 → 256"]
    Classifier --> Output["Output<br/>FAKE/REAL<br/>+ confidence"]

    style Input fill:#1e293b,stroke:#3b82f6,stroke-width:2px,color:#fff
    style Spatial fill:#0f172a,stroke:#10b981,stroke-width:2px,color:#10b981
    style Frequency fill:#0f172a,stroke:#3b82f6,stroke-width:2px,color:#3b82f6
    style FFT fill:#0d1117,stroke:#06b6d4,color:#06b6d4
    style DCT fill:#0d1117,stroke:#06b6d4,color:#06b6d4
    style Fusion fill:#0d1117,stroke:#3b82f6,color:#3b82f6
    style Classifier fill:#0f172a,stroke:#ec4899,stroke-width:2px,color:#ec4899
    style Output fill:#1e293b,stroke:#a855f7,stroke-width:2px,color:#fff
    style Concat fill:#0d1117,stroke:#8b5cf6,color:#8b5cf6
    style FFT_Out fill:#0d1117,stroke:#64748b,color:#64748b
    style DCT_Out fill:#0d1117,stroke:#64748b,color:#64748b
    style Spatial_Out fill:#0d1117,stroke:#64748b,color:#64748b
    style Fusion_Out fill:#0d1117,stroke:#64748b,color:#64748b

Loading

Model Specifications

Specification	Value
Total Parameters	25.05M params
Input Size	224×224 RGB
Output	Binary (FAKE/REAL) + confidence
Backbone	EfficientNet-B4 (19.34M params)
Frequency Module	2.16M params
Classifier	3.54M params

---

Training

Infrastructure We trained the model on an RTX 3090 (with CUDA) for approximately 4 hours. We used the GPU provider vast.ai The training file is located in the training/trainer.py module

Specs GPU Framework Duration

---

Dataset

Source We started from an existing dataset found on Kaggle: FaceForensics++ Dataset (C23) Containing 7000 videos with numerous deepfake techniques: We extracted the frames and faces from these videos to create our dataset: VeridisQuo Preprocessed Dataset Preprocessing Pipeline The dataset was built using the following pipeline: Frame Extraction: 1 FPS from videos (PyAV GPU-accelerated) Face Detection: YOLOv11n-face-detection (confidence ≥ 0.7) Face Extraction: 224×224 crops with 20px padding Dataset Split: Stratified 70/15/15 split Class Balancing: Oversample minority class

Distribution Split Samples Ratio Train 499,965 70% Test 107,620 15% Eval 108,853 15% Total: 716,438 images

---

Configuration & Results

Configuration Parameter Value batch_size 64 learning_rate 0.0001 min_learning_rate 0.000001 num_epochs 7 weight_decay 0.0001 optimizer AdamW scheduler Warmup + Cosine Annealing warmup_epochs 2 use_automixed_precision false loss_func CrossEntropyLoss

Results Training Accuracy Training Loss

API Reference

Endpoints

Method	Endpoint	Description
GET	/api/v1/health	Health check and model status
POST	/api/v1/analyze	Analyze video for deepfakes
GET	/api/v1/outputs/{filename}	Download GradCAM visualization
DELETE	/api/v1/outputs/{filename}	Delete output file

Request Format (POST /api/v1/analyze)

curl -X POST http://localhost:8000/api/v1/analyze \
  -F "file=@video.mp4" \
  -F "fps=1" \
  -F "aggregation_method=majority" \
  -F "generate_gradcam=true"

Parameters:

• file: Video file (MP4, AVI, MOV, MKV, WEBM)
• fps: Frames per second to extract (default: 1)
• aggregation_method: Score aggregation (default: majority)
• generate_gradcam: Generate visualization video (default: false)

Response Format

{
  "prediction": "FAKE",
  "confidence": 0.8734,
  "aggregation_method": "majority",
  "total_frames": 120,
  "gradcam_video_path": "/api/v1/outputs/gradcam_video_20250102.mp4"
}

---

Quick Start

Prerequisites

• Python 3.12 or 3.13
• uv package manager
• Node.js 18+ and npm (optional, for frontend)
• CUDA 11.8+ (optional, for GPU acceleration)

Clone project

# Clone repository
git clone https://github.com/VeridisQuo-orga/VeridisQuo.git
cd VeridisQuo

Launch backend

chmod +x ./scripts/launch_api.sh
./scripts/launch_api.sh

server runs on http://localhost:8000 | Docs at /docs

Launch frontend

chmod +x ./scripts/launch_frontend.sh
./scripts/launch_frontend.sh

Development server on http://localhost:3000

---

Citation

If you use VeridisQuo in your research, please cite:

@software{veridisquo2025,
  title = {VeridisQuo: Hybrid Deepfake Detection with Explainable AI},
  author = {Castillo, Theo and Barriere, Clement},
  year = {2025},
  url = {https://github.com/VeridisQuo-orga/VeridisQuo},
  note = {Model: \url{https://huggingface.co/Gazeux33/VeridisQuo}}
}