alihaidar-ai/aegis-ai-multimedia-integrity-full
🛡️ AegisAI — Multimedia Integrity Analyzer (Video + Photo)
AegisAI is a Streamlit-based multimedia integrity and deepfake risk analyzer that evaluates suspect videos and photos for potential manipulation (deepfake, face-swap, or heavy edits) using a combination of:
Quality signals (blur + noise → baseline risk)
Multi-face extraction (video frames + photos)
Model-based deepfake scoring
Optional identity verification (Reference vs Suspect)
Evidence strength scoring
Downloadable forensic reports (CSV)
⚠️ Disclaimer: This tool provides an AI-assisted assessment, not legal proof. Results may be incorrect. Always verify with human experts and additional forensic methods.
🚀 Live Demo
👉 Hugging Face Space: https://huggingface.co/spaces/alihaidar-ai/aegis-ai-multimedia-integrity-full
👉 Streamlit Cloud: https://aegis-ai-multimedia-integrity-4vypudbpyourhrkul5xu7h.streamlit.app/
✨ Features 🎞️ Suspect VIDEO Analysis
Upload suspect video (mp4, mov, avi)
Extract frames with slider control
Compute:
Blur score
Noise score
Baseline risk (0–100)
Detect faces and display 224×224 face crops
Multi-person clustering (if identity enabled)
Download:
Video frames evidence CSV
Deepfake per-face scores CSV
🖼️ Suspect PHOTO Analysis
Upload one or more suspect photos (jpg, jpeg, png)
Detect faces and show 224×224 crops
Optional multi-person check
Download deepfake per-face scores CSV
🪪 Optional Identity Verification
Upload reference photos (1–5 clear selfies)
Upload reference videos (1–3 short clips)
Builds identity embedding
Filters inconsistent reference media
Compares suspect faces vs reference
Outputs:
MATCH
UNCERTAIN
NOT MATCH
📊 Final Risk & Smart Verdict
Combines:
Deepfake probability
Baseline risk
Identity similarity (if provided)
Produces:
Final Risk Score (0–100)
Evidence Strength Score (0–100)
Smart Verdict Summary
📥 Downloadable Report
CSV report includes:
final_risk
evidence_strength
identitysimilaritybest
fakescorebest
avg_blur
avgbaselinerisk
identity_enabled
reference_provided
video_uploaded
numvideofaces
numphotofaces
🧠 Tech Stack
Python
Streamlit
OpenCV
NumPy
Pandas
Pillow
Hugging Face Transformers
InsightFace
📁 Project Structure . ├── app.py ├── streamlitapp.py ├── utilsvideo.py ├── utilsidentity.py ├── utilsdeepfake_model.py ├── requirements.txt └── README.md 🛠️ Local Setup 1️⃣ Create virtual environment python -m venv venv
Windows:
venv\Scripts\activate
Linux/Mac:
source venv/bin/activate 2️⃣ Install dependencies pip install -r requirements.txt 3️⃣ Run the app streamlit run app.py ⚠️ Notes & Limitations
Accuracy depends on media quality.
Small or occluded faces may fail detection.
Deepfake models may produce false positives/negatives.
Identity verification accuracy depends on clean reference media.
This is an assistive AI tool, not a forensic authority.
👤 Author
Ali Haidar BS Computer Science (2020–2024) AI / ML Projects Portfolio
GitHub: https://github.com/studyhaidar77-art
⭐ Support
If you find this project useful, please ⭐ the repository.