An advanced, production-ready AI forensic platform designed to combat highly sophisticated generative media, including AI-generated images, deepfake videos, voice cloning, and synthetically manipulated social media links.
The system leverages a multi-microservice architecture to process massive spatial, temporal, and spectral tensor payloads concurrently, unifying them via a deterministic Cross-Modal Consensus Hub to detect zero-day digital manipulation.
A dedicated, isolated PyTorch microservice executing a 5-layer analysis pipeline:
ai_detector_best.pth): Evaluates macro-structural artifacts caused by generative scaling.manipulation_detector_best.pth): Scans the stacked ELA tensor for splicing anomalies.ai_detector_best & manipulation_detector_best).wav2vec2-base-960h acoustic transformer.calculate_lipsync_deviation): Mathematically traces visual kinetic text boundaries versus auditory spectral spikes.verify_semantic_alignment): Ingests and enforces prompt consistency via mathematically calculated latent alignment algorithms (Cosine Similarity).To avoid the network latency of cloud APIs and ensure absolute privacy, the system relies on proprietary local .pth neural weight checkpoints heavily fine-tuned specifically for this platform.
ai_detector_best.pth (EfficientNet-B0 Baseline):
manipulation_detector_best.pth (ResNet-50 Patch Analyzer):
Nexus-AudioForge-v2.pth (wav2vec2-base-960h Transformer):
This system employs a decoupled, multi-tier architecture to securely parallelize heavy tensor operations.
framer-motion for complex continuous 50-second terminal logging telemetry./models): All custom trained .pth topologies are localized on-device to ensure privacy and circumvent external API latency.db.json logging (SQLite-scalable)Ensure you have the following installed on your machine:
git clone https://github.com/your-org/multimodal-fake-detector.git
cd multimodal-fake-detector
# Create and activate virtual environment
python -m venv venv
# Windows:
.\venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate
# Install strictly enforced AI dependencies
pip install -r requirements.txt
Ensure your custom trained weights are located in the unified ./models directory:
models/ai_detector_best.pthmodels/manipulation_detector_best.pthcd frontend
npm install
cd ..
To achieve fully optimized local inference, you will need to start three distinct services simultaneously.
This server handles incoming forensic streams, the Audio pipeline, the Video extraction logic, and multi-fusion data aggregation.
.\venv\Scripts\activate
uvicorn api:app --host 0.0.0.0 --port 8000 --reload
This specialized microservice spins up PyTorch solely to process the ELA Spatial pipeline safely outside the main thread.
.\venv\Scripts\activate
uvicorn image_analysis.api.main:app --host 0.0.0.0 --port 8002 --reload
Initializes the React user interface.
cd frontend
npm run dev
The application is now accessible locally. Proceed to http://localhost:5173.
Proprietary Forensic Software. Use for academic or verified hackathon/demonstration purposes only. Do not scale for public production without establishing load-balancing on the internal PyTorch tensor allocation clusters.
An advanced, production-ready AI forensic platform designed to combat highly sophisticated generative media, including AI-generated images, deepfake videos, voice cloning, and synthetically manipulated social media links.
The system leverages a multi-microservice architecture to process massive spatial, temporal, and spectral tensor payloads concurrently, unifying them via a deterministic Cross-Modal Consensus Hub to detect zero-day digital manipulation.
A dedicated, isolated PyTorch microservice executing a 5-layer analysis pipeline:
ai_detector_best.pth): Evaluates macro-structural artifacts caused by generative scaling.manipulation_detector_best.pth): Scans the stacked ELA tensor for splicing anomalies.ai_detector_best & manipulation_detector_best).wav2vec2-base-960h acoustic transformer.calculate_lipsync_deviation): Mathematically traces visual kinetic text boundaries versus auditory spectral spikes.verify_semantic_alignment): Ingests and enforces prompt consistency via mathematically calculated latent alignment algorithms (Cosine Similarity).To avoid the network latency of cloud APIs and ensure absolute privacy, the system relies on proprietary local .pth neural weight checkpoints heavily fine-tuned specifically for this platform.
ai_detector_best.pth (EfficientNet-B0 Baseline):
manipulation_detector_best.pth (ResNet-50 Patch Analyzer):
Nexus-AudioForge-v2.pth (wav2vec2-base-960h Transformer):
This system employs a decoupled, multi-tier architecture to securely parallelize heavy tensor operations.
framer-motion for complex continuous 50-second terminal logging telemetry./models): All custom trained .pth topologies are localized on-device to ensure privacy and circumvent external API latency.db.json logging (SQLite-scalable)Ensure you have the following installed on your machine:
git clone https://github.com/your-org/multimodal-fake-detector.git
cd multimodal-fake-detector
# Create and activate virtual environment
python -m venv venv
# Windows:
.\venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate
# Install strictly enforced AI dependencies
pip install -r requirements.txt
Ensure your custom trained weights are located in the unified ./models directory:
models/ai_detector_best.pthmodels/manipulation_detector_best.pthcd frontend
npm install
cd ..
To achieve fully optimized local inference, you will need to start three distinct services simultaneously.
This server handles incoming forensic streams, the Audio pipeline, the Video extraction logic, and multi-fusion data aggregation.
.\venv\Scripts\activate
uvicorn api:app --host 0.0.0.0 --port 8000 --reload
This specialized microservice spins up PyTorch solely to process the ELA Spatial pipeline safely outside the main thread.
.\venv\Scripts\activate
uvicorn image_analysis.api.main:app --host 0.0.0.0 --port 8002 --reload
Initializes the React user interface.
cd frontend
npm run dev
The application is now accessible locally. Proceed to http://localhost:5173.
Proprietary Forensic Software. Use for academic or verified hackathon/demonstration purposes only. Do not scale for public production without establishing load-balancing on the internal PyTorch tensor allocation clusters.