📝 Project Documentation: Brillio Fake Review Detection System

GitHub Links 📂:

• 📱 iOS: AR-iOS-APP
• 🌐 Chrome Extension: Chrome-Plugin
• 🤖 PyTorch Model: AI-Python

Our project focuses on creating a robust and versatile fake review detection system powered by an AI model. The solution spans multiple platforms 🌍 and integrates different technologies, including a machine learning model designed to detect fake reviews, an iOS app built to scan barcodes and authenticate product reviews 📱, and a Chrome extension to provide on-device fake review detection 🌐.

We began by designing and training an AI model specifically built to identify fake product reviews. This model uses an LSTM neural network 🧬, which is ideal for processing text-based data. Here’s a breakdown:

• 📊 Data Collection: We collected a labeled dataset, categorizing each review as "OR" (legitimate) or "CG" (fake). This dataset, stored in a CSV file, provided the foundation for model training.
• 🛠️ Preprocessing: Using NLTK for text preprocessing, we tokenized, removed stop words, and vectorized text to ensure optimal training.
• 🔗 Model Architecture: The PyTorch-based LSTM model has an embedding layer, bidirectional LSTM layers, and fully connected layers for advanced contextual understanding of reviews.
• 📈 Training and Evaluation: Through iterative training, we fine-tuned the model’s performance, achieving over 90% accuracy in detecting fake reviews. ⚙️

Challenges:

• 📝 Dealing with batch sizes larger than 1 during ONNX export generated warnings. We optimized the model export to be compatible with a batch size of 1.

To integrate the AI model into an iOS environment, we converted it to Core ML format, allowing on-device inference within Apple’s ecosystem 🍏.

• 🔄 Conversion Process: We used torchvision and coremltools to transform the model, adding configuration for probability-based outputs and class labels.
• 🚀 Performance Optimization: The conversion to Core ML ensured low latency and quick responses, ideal for mobile devices.

Challenges:

• ⚠️ Core ML compatibility with LSTM layers required adjustments in input handling and metadata to ensure smooth functioning.

For a real-world application, we developed an iOS app that lets users scan barcodes 📱 and verify review authenticity.

• 📐 Architecture: SwiftUI delivers a responsive interface, and AVFoundation powers the barcode scanning feature.
• 🤖 Core ML Integration: The Core ML model analyzes reviews fetched from a database, providing real-time feedback.
• 🖥️ User Experience: Color-coded indicators display review authenticity for quick user assessment.

Challenges:

• Ensuring smooth Core ML processing for real-time feedback without impacting performance was essential.

We also developed a Chrome extension 🌐 that allows users to identify fake reviews directly from their browser.

• ⚙️ ONNX Model Deployment: To maintain cross-platform compatibility, we used an ONNX model that runs entirely on-device for user privacy.
• 🌐 Extension Workflow: The extension analyzes reviews as users browse, offering real-time insights.
• 🖱️ User Interaction: Visual feedback provides clear indicators of review authenticity.

Challenges:

• Ensuring lightweight model performance within the Chrome environment required optimizations, especially compared to Core ML’s mobile framework.

Throughout development, we encountered several key challenges:

• 🖥️ Model Compatibility Across Platforms: Adapting the model for Core ML and ONNX required testing to ensure seamless performance.
• 🕒 Performance Optimization: Real-time feedback, especially on mobile and in-browser, demanded model size reductions and hyperparameter fine-tuning.
• 🔒 User Privacy: To protect privacy, both the iOS app and Chrome extension function entirely offline.

This project showcases an effective AI solution that enhances decision-making through review authenticity insights. With iOS and web integration, our system is versatile, privacy-focused, and user-friendly. Overcoming various technical challenges strengthened our skills in machine learning, cross-platform model deployment, and user-centered development 🚀.