A hybrid deep learning system to detect deepfakes in images and videos using EfficientNetB4 and CNN-GRU, combined with facial emotion analysis using DeepFace. Designed to be fast, scalable, and suitable for real-world deployment.

• ✅ Detects deepfakes in both images and videos
• 📊 Generates real-time metrics: Accuracy, Precision, Recall, F1, AUC
• 🎭 Performs emotion recognition on detected faces using DeepFace
• ⚙️ Uses EfficientNetB4 for feature extraction
• 🧠 Includes CNN + GRU model for temporal video analysis
• 📈 Live plots for deepfake probability and emotion timeline
• 💻 Lightweight, efficient & optimized for low-resource environments

.
├── models/
│   ├── efficientnetB4_deepfake_final.keras
│   ├── cnn_gru_model.h5
│
├── real_and_fake_face/
│   ├── real/
│   └── fake/
│
├── real and fake video/
│   ├── real/
│   └── fake/
│
├── test real and fake video/
│   ├── real/
│   └── fake/
│
├── deepfake_train.py                   # Train EfficientNetB4 image model
├── evaluate_image_model.py            # Evaluate on static images
├── analyze_single_image.py            # Image deepfake + emotion detection
├── analyze_video_with_emotion.py      # Video deepfake + emotion + graph
├── cnn_gru_train.py                   # Train CNN + GRU on video frames
├── cnn_gru_evaluation.py              # Evaluate videos & log predictions

EfficientNetB4 (frozen)
→ GlobalAveragePooling2D
→ Dropout + Dense
→ Sigmoid output (real vs fake)

TimeDistributed(EfficientNetB4)
→ GlobalAveragePooling2D (per frame)
→ GRU layer
→ Dropout + Dense
→ Sigmoid output

• Performed using DeepFace
• Extracted directly from video frames or static images
• Tracked over time in videos with frame-wise stats and plots

🎭 Emotions supported: happy, sad, angry, surprise, neutral, etc.

• Deepfake Probability Timeline across video frames

• Emotion Timeline Plot to show mood shifts over time

• Saved as:

  • deepfake_video_graph.png
  • emotion_over_time_graph.png

python deepfake_train.py

python evaluate_image_model.py

python analyze_single_image.py

python analyze_video_with_emotion.py

python cnn_gru_train.py

python cnn_gru_evaluation.py

• real_and_fake_face/ — manually balanced dataset with real & fake face images
• real and fake video/ — training videos categorized as real/fake
• test real and fake video/ — separate test set for generalization

Data preprocessing includes:

• Image validation and cleaning
• Frame extraction with resizing
• Haar cascade face detection

• CNN-GRU video model improves temporal understanding vs. static image detection.

• Emotion analysis adds context to deepfake results.

• Decision logic includes:

  • Average probability
  • Max probability
  • Smoothed curve thresholding

Frame 120 | Deepfake Probability: 0.8723 | Emotion: neutral  
Frame 125 | Deepfake Probability: 0.9231 | Emotion: happy  

Average Deepfake Probability: 0.89  
Likely: 🚨 Deepfake  

• Add race, age, and gender analysis using DeepFace
• Deploy model using Streamlit or Flask Web App
• Use Hugging Face Hub for model sharing
• Add real-time webcam streaming analysis

Harshit Bansal 📫 [email protected] 🔗 GitHub | LinkedIn

This project is licensed under the MIT License.
Feel free to use, adapt, and cite the work with attribution.