RJPOLICE_HACK_1400_BitwiseOperators_8

Deepfake Audio and AI Content Detection

An integrated project combining multiple repositories to detect deepfake audio, AI-generated images/videos, and AI-generated text using deep learning, signal processing, and face analysis techniques.

Built for research, awareness, and educational purposes to counteract the rising threat of deepfake and synthetic media.

---

Features

• Audio Deepfake Detection — identifies synthetic speech from mel-spectrogram features using a trained Keras CNN (weights/Deepfake_audio.h5).
• Image & Video Deepfake Detection — MesoInception-4 and ResNet CNN models detect manipulated faces (weights/MesoInception_F2F.h5).
• AI-Generated Text Detection — a BERT sequence classifier (served via the Flask app) estimates the probability that a passage was written by an AI.
• Face Extraction — extractFaceFromVideo.py uses MTCNN to crop faces from video frames for downstream analysis.
• GUI Tools — Tkinter desktop apps and a Flask web app to load media, run analysis, and read results.
• Batch CLI (new) — detect_audio.py, a headless command-line tool that scores one file or a whole folder of audio and exports JSON/CSV reports.

---

Directory Structure

.
├── Src/
│   ├── app.py              # Flask web app (audio deepfake + AI-text detection)
│   ├── templates/          # index.html, result.html
│   └── uploads/            # runtime upload folder
├── weights/                # Pre-trained model weights (.h5)
│   ├── Deepfake_audio.h5
│   └── MesoInception_F2F.h5
├── ScreenShots/            # Example outputs and GUI visuals
├── detect_audio.py         # NEW: headless batch audio classifier (CLI)
├── final_gui.py            # Tkinter GUI: audio + MesoInception image detection
├── resnetgui.py            # Tkinter GUI: ResNet image classifier
├── partialGUI.py           # Tkinter chat-style detection UI prototype
├── extractFaceFromVideo.py # MTCNN face extraction from video
├── *.ipynb                 # Training/experiment notebooks for each model
└── requirements.txt        # Python dependencies

---

Getting Started

Installation

git clone https://github.com/Developer1010x/Deepfake-Audio-and-AI-Content-Detection.git
cd Deepfake-Audio-and-AI-Content-Detection
pip install -r requirements.txt

Run the Flask web app

cd Src
python app.py
# then open http://127.0.0.1:5000

The web app expects Deepfake_audio.h5 for audio and a BERT model/ directory for the text classifier.

Run a desktop GUI

python final_gui.py     # audio + image (MesoInception) classifier
python resnetgui.py     # ResNet image classifier

Note: the GUI scripts contain hard-coded Windows-style model paths (e.g. D:\Desktop\hack\...). Update those paths to point at the files under weights/ before running.

---

Batch Audio Detection (CLI)

detect_audio.py runs the audio deepfake model without any GUI or web server, reusing the exact same mel-spectrogram preprocessing as the rest of the project. It accepts a single file or a directory and can emit JSON/CSV reports.

# Score a single file (uses weights/Deepfake_audio.h5 by default)
python detect_audio.py sample.wav

# Recursively score every audio file in a folder and write a CSV report
python detect_audio.py ./clips --csv results.csv

# Point at a different model and print JSON to stdout
python detect_audio.py sample.mp3 --model weights/Deepfake_audio.h5 --json -

# Quiet mode (suppress per-file lines), JSON to file
python detect_audio.py ./clips --quiet --json results.json

Supported formats: .wav, .mp3, .flac, .ogg, .m4a.

Example output:

    Real  ( 92.4%)  clips/genuine_01.wav
Deepfake  ( 88.1%)  clips/synth_02.wav

Exit codes: 0 success, 1 bad input / nothing classified, 2 model load failure.

---

Models & Weights

Task	Model	Weights file
Audio deepfake	Keras CNN	weights/Deepfake_audio.h5
Image deepfake	MesoInception-4	weights/MesoInception_F2F.h5
Image deepfake (alt)	ResNet	resnet.h5 (not bundled)
AI-text detection	BERT classifier	model/ (not bundled)

---

Disclaimer

These models are provided for educational and research use. Predictions are probabilistic and may be wrong; do not rely on them as sole evidence for any high-stakes decision.