# 🧠 Deepfake Detection with Emotion Analysis using EfficientNet and DeepFace A hybrid deep learning system designed to detect **deepfakes** in both images and videos, while simultaneously performing **facial emotion analysis**. Built using EfficientNetB4, CNN-GRU, and DeepFace, the system is lightweight, scalable, and tailored for real-world deployment. --- ## 🚀 Key Features - 🔍 Detects deepfakes in **images and videos** - 🧠 Performs **emotion recognition** using DeepFace - 📊 Live plotting of **deepfake probability** and **emotion timeline** - 🧮 Metrics: Accuracy, Precision, Recall, F1, AUC - ⚙️ EfficientNetB4 + CNN-GRU architecture for temporal video analysis - 💻 Optimized for **low-resource systems** --- ## 🧰 Tech Stack | Area | Tools & Libraries | | ---------------- | ------------------------------------------ | | Deep Learning | TensorFlow, Keras, EfficientNetB4, GRU | | Face Analysis | DeepFace (for emotion detection) | | Image Processing | OpenCV, Haar Cascades | | Data | NumPy, pandas, scikit-learn | | Visualization | Matplotlib | --- ## 🧠 Model Architectures ### 🔹 EfficientNetB4 (Image-Based) ``` EfficientNetB4 (frozen) → GlobalAveragePooling2D → Dropout + Dense → Sigmoid output (real vs fake)
TimeDistributed(EfficientNetB4) → GlobalAveragePooling2D (per frame) → GRU → Dense → Sigmoid output
deepfake_train.py — Train image-based modelevaluate_image_model.py — Evaluate on static imagesanalyze_single_image.py — Image + emotion detectionanalyze_video_with_emotion.py — Video analysis with live plotscnn_gru_train.py — Temporal model training (video)cnn_gru_evaluation.py — CNN-GRU video evaluation| Model | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|
| EfficientNetB4 (Base) | 48.1% | 45.6% | 53.5% | 49.2% |
| Fine-tuned Model | 57.7% | 53.6% | 74.6% | 62.3% |
happy, sad, angry, surprise, neutral, etc.deepfake_video_graph.png — Deepfake probability over time
emotion_over_time_graph.png — Emotion timeline from video
Combined real-time predictions:
Frame 120 | Deepfake: 0.8723 | Emotion: neutral Frame 125 | Deepfake: 0.9231 | Emotion: happy
Real and Fake Face Detection:
https://www.kaggle.com/datasets/ciplab/real-and-fake-face-detection
Deepfake Detection on Images and Videos (video-based):
https://www.kaggle.com/code/krooz0/deep-fake-detection-on-images-and-videos/input
Data processing steps include:
# Train EfficientNetB4 python deepfake_train.py # Evaluate image model python evaluate_image_model.py # Analyze single image (deepfake + emotion) python analyze_single_image.py # Analyze video (deepfake + emotion timeline) python analyze_video_with_emotion.py # Train temporal model (CNN + GRU) python cnn_gru_train.py # Evaluate temporal model python cnn_gru_evaluation.py
Harshit Bansal
📫 harshitbansalmi@gmail.com
🔗 GitHub | LinkedIn
Licensed under the MIT License.
Free to use, adapt, and cite with attribution.