🎧 Multi‑module Deepfake Audio Detector

📝 Project Type: Research project
📚 Course: Digital Forensics

---

Project Overview

This repository contains a Jupyter Notebook implementation of a multi-module deepfake audio detection system. The detector analyzes audio samples to identify manipulated (deepfake) content using a combination of feature extraction, machine learning models, and decision fusion techniques.

---

🎯 Motivation & Forensics Impact

Deepfake audio—synthetic or manipulated voice recordings—pose a growing threat in the forensics domain:

• Trusted Evidence at Risk: Audio recordings often serve as key evidence in investigations. Undetected deepfakes can mislead courts or investigations.
• Privacy & Security Threats: Malicious actors may impersonate individuals to deceive victims or manipulate public opinion.
• Forensic Integrity: Robust detection tools are critical to maintain trust in audio-based evidence and uphold justice in legal systems.

---

🛠️ Techniques Used

Module	Description
Entropy-Based	Logistic Regression on entropy-based features
MFCC-Based	Random Forest using Mel-Frequency Cepstral Coefficients
MelCNN	CNN trained on mel-spectrograms
RawWaveformCNN	CNN trained directly on waveform inputs

Each model is evaluated individually, then combined using ensemble decision strategies for better generalization and detection reliability.

---

📈 Experimental Results

After training and evaluating on the ASVspoof 2019 dataset:

• Entropy Classifier (Logistic Regression)

  • Validation Accuracy: ~81.2%
  • Test Accuracy: ~80.5%

• MFCC Classifier (Random Forest)

  • Validation Accuracy: ~92.6%
  • Test Accuracy: ~90.7%

• MelCNN

  • Validation Accuracy: ~90.3%
  • Test Accuracy: ~89.5%

• RawCNN

  • Validation Accuracy: ~91.1%
  • Test Accuracy: ~89.4%

---

🧪 Dataset

• Dataset used: ASVspoof 2019 LA (Logical Access subset)
• Format: FLAC audio files
• Labels: bonafide (real), spoof (deepfake)

📌 Dataset must be pre-downloaded and path configured in main() function.

---

🧩 System Architecture

This multi-module approach enhances detection accuracy and robustness:

1. Feature Extraction

   • Extract acoustic and spectral features such as MFCCs, pitch, and spectral centroid.

2. Model Ensemble

   • Train multiple detectors (e.g., SVM, Random Forest, CNN) on the extracted features.

3. Decision Fusion

   • Combine model outputs using voting or weighted averaging to reach a final classification (genuine vs. deepfake).

---

🛠️ Usage Instructions

1. Ensure required libraries are installed:

 pip install -r requirements.txt

2. Launch the notebook:

   jupyter notebook Multi_module_Deepfake_Audio_Detector.ipynb

3. Follow these steps within the notebook:

   • Load your audio dataset (genuine vs. deepfake)
   • Run feature extraction
   • Train and evaluate individual models
   • Apply decision fusion to combine model predictions
   • Analyze results: accuracy, precision, recall, ROC curves

---

📂 File Structure

├── Multi\_module\_Deepfake\_Audio\_Detector.ipynb   # Main notebook
├── features\_entropy.py                          # Entropy feature extractor
├── features\_mfcc.py                             # MFCC extractor
├── features\_melcnn.py                           # MelCNN model + extractor
├── features\_rawcnn.py                           # Raw waveform CNN
└── utils, models, configs 

---

Requirements

• Python 3.7+

• Jupyter Notebook

• Key libraries:

  numpy, pandas, librosa, scikit-learn, tensorflow (or torch), matplotlib, seaborn
  

---

Forensic Use & Relevance

This tool provides digital forensic analysts with a practical, modular detection pipeline to verify audio authenticity. It underlines the importance of:

• Reliable validation of audio-based evidence
• Scientific rigour through multi-model ensemble techniques
• Transparency, enabling traceability of detection steps—crucial in legal contexts

---

🎓 Academic Context

This project was completed for the Digital Forensics course as part of the BSc Cybersecurity program. It aims to explore the applicability of machine learning to deepfake detection in modern forensics.