🔊 DeepFake Voice Detection

A robust machine learning system for identifying AI-generated (deepfake) voices using MFCC features, spectral audio statistics, and a combination of Deep Learning (CNN) and Classical ML (Random Forest, KNN) models.

---

⭐ Key Features

• Detects real vs fake speech with 98%+ accuracy
• Uses MFCCs, Mel-spectrogram statistics, and spectral features
• Implements 1D CNN, Random Forest, and KNN
• Includes full evaluation metrics with placeholders for all plots
• Built to support future deployment for fraud and security applications

---

📂 Dataset

• Audio format: 16 kHz, mono WAV
• Classes: Real, Fake
• Includes various SNR levels and noise-reduction methods
• Dataset contains class imbalance, handled using SMOTE

---

🎧 Feature Extraction

MFCC Features (40-D)

• Extracted using Librosa
• Mean-pooled across time
• Used as input to the 1D CNN

Engineered Audio Features (318-D)

• MFCC means
• Mel-Spectrogram means
• Log-Spectrogram (STFT) means
• Used for Random Forest and KNN models

---

🏗️ System Architecture

           ┌─────────────── Preprocessing ───────────────┐
           │                                               │
Audio → Load → Normalize → MFCC / Spectrogram Extraction → Features
           │                                               │
           └──────────────────────┬────────────────────────┘
                                  │
       ┌──────────────────────────┴──────────────────────────┐
       │                                                     │
  318-D Engineered Features                          MFCC Map (40×1)
       │                                                     │
Random Forest / KNN                                     1D CNN Model
       │                                                     │
       └──────────────────────────┬──────────────────────────┘
                                  ↓
                           **Real / Fake**

---

🧠 Models Implemented

1️⃣ 1D CNN (MFCC-Based)

✔ Architecture

• Conv1D → Dropout
• MaxPooling
• Conv1D → Dropout
• Dense + Softmax
• Trained for 40 epochs

✔ Features Used

✔ Performance

• Dev Accuracy: ~86%
• Eval Accuracy: ~88%
• Strong on detecting fake audio
• Mild overfitting

---

📊 CNN Evaluation (Placeholders)

🟦 Training Accuracy Curve

🟥 Training Loss Curve

🟩 Confusion Matrix (CNN)

🟪 ROC Curve (CNN)

🟨 Precision–Recall Curve (CNN)

---

✔ Features Used For the 318 Dimention

2️⃣ Random Forest Classifier

✔ Performance

• Accuracy: 98.82%
• High precision & recall on both classes
• Extremely robust to noise & dataset variance

🟩 Confusion Matrix (RF)

---

3️⃣ K-Nearest Neighbours (KNN)

✔ Performance

• Accuracy: 98.29%
• Very stable across different samples
• k = 7 chosen for optimal performance

🟩 Confusion Matrix (KNN)

---

📊 Overall Model Comparison

Model	Accuracy	Strengths	Weaknesses
Random Forest	⭐ 98.82%	Best overall, robust to noise	Slow to train on huge datasets
KNN (k=7)	98.29%	Simple & competitive	Slow inference on large data
CNN (MFCCs)	~88%	Learns temporal patterns	Overfitting risk

---

🚧 Limitations

• Dataset imbalance required oversampling
• CNN performance limited by MFCC-only representation
• Needs evaluation on unseen deepfake generators
• Real-world recordings with background noise not fully tested

---

🚀 Future Enhancements

• Use 2D CNNs on spectrogram images
• Add transformer-based encoders (wav2vec 2.0, HuBERT, Whisper)
• Deploy as a web or mobile app for live detection
• Add adversarial robustness
• Add explainable AI for forensic usage

---

🛠️ Tech Stack

• Python
• Librosa – audio processing
• TensorFlow / Keras – CNN model
• scikit-learn – RF, KNN, SMOTE
• NumPy / pandas – preprocessing

---

🙌 Contributors

Srujan Rana

---