SAVANA_API

SAVANA: Secure AI Verification And Non-deceptive Analysis

SAVANA API, encompassing four key features: image deepfake detection, video deepfake detection, AI content detection in research papers, and AI content detection in blogs. Utilizing advanced machine learning models and natural language processing techniques, our system ensures reliable and accurate identification of AI-generated content, thus enhancing the integrity and trustworthiness of digital media. The proposed methodology includes the use of BlazeFace for face detection, EfficientNetB4 for regression tasks, fine-tuned vectorizers with logistic regression for text analysis, and LLaMA 3 for nuanced linguistic pattern detection. This study outlines the technical aspects and experimental results of each feature, demonstrating the efficacy of our approach in combating deceptive AI content.

Abstract

This project presents the design and implementation of a comprehensive system for secure AI verification and non-deceptive analysis. Our system encompasses four key features:

1. Image Deepfake Detection
2. Video Deepfake Detection
3. AI Content Detection
4. AI Content Detection in Research Papers
5. AI Content Detection in Blogs -This README outlines the technical aspects and provides an overview of the experimental results for each feature, demonstrating the efficacy of our approach in combating deceptive AI content.

Demo:

Flowchart:

Features

1. Image Deepfake Detection

Detects manipulated images using BlazeFace for face detection and EfficientNetB4 for the regression tasks.

2. Video Deepfake Detection

Analyzes videos to identify deepfake content by leveraging advanced machine learning techniques.

3. AI Content Detection in Research Papers

Employs fine-tuned vectorizers and logistic regression to detect AI-generated content in academic research papers.

4. AI Content Detection in Blogs

Uses LLaMA 3 to analyze and identify AI-generated content in blogs, ensuring the authenticity of the text.

Requirements

• tensorflow
• keras
• numpy
• opencv-python
• blazeface
• efficientnet

text detection libraries:

• os
• sys
• csv
• joblib
• logging
• nltk
• ssl
• spacy
• python-docx
• PyPDF2
• textblob
• scikit-learn

Downloads

• BlazeFace Model: BlazeFace GitHub
• EfficientNetB4 Weights: TensorFlow Hub
• DFDC Dataset: Kaggle

(in text detection code)

• import nltk nltk.download('punkt') nltk.download('stopwords')

model.pkl

This file contains the trained logistic regression model saved using joblib.

vectorizer.pkl

This file contains the TF-IDF vectorizer used to transform the text data, saved using joblib.

Usage

Detect Deepfakes in Images

bash python detect_deepfake.py --input image --path /path/to/image.jpg

Detect Deepfakes in Videos

bash python detect_deepfake.py --input video --path /path/to/video.mp4

Training

To train the EfficientNetB4 model:

bash python train_model.py --data /path/to/dfdc_dataset

Installation Steps

1. Clone the Repository:

   git clone https://github.com/techcodebhavesh/SAVANA_API.git
   

2. Navigate to Project Directory:

   cd SAVANA_API
   

3. Install Dependencies:

   • To ensure all dependencies are installed correctly, navigate to the project's client and root directory and run the following command:

   npm install
   

4. Fix for Legacy Peer Dependencies:

   • If you encounter any issues related to legacy peer dependencies during the installation process, use the following command to install them:

   npm i --legacy-peer-deps
   

5. Build the Project:

   npm run build
   

6. Download and Set up the Llama3 model on the Linux server: -run the server code

   node ser.js
   

7. Start the Python Server:

   • Within the project's respective directory(/AI/setup/deepfake/deep), execute the command:

   python runpython.py
   

8. Start the Application:

   • Within the project's main directory, execute the command:

   npm run start