🛡️ EmpowerNet AI

Multi-Modal Deepfake & Scam Detection Platform

EmpowerNet AI is an explainable, multi-modal security system that detects scams, phishing, social engineering, and AI-generated deepfakes across text, images, audio, and video.
It combines transformer models, computer vision, signal forensics, and heuristic risk scoring to deliver real-time, trustworthy threat analysis.

DEMO VIDEO-https://drive.google.com/file/d/1N0-l76tcmqyVLoJneCnO7WZ6s42YjIZt/view?usp=sharing

PPT-https://drive.google.com/file/d/1ZtSXO3dTaolsT6nSjgyIdRULqXTgTIYL/view?usp=sharing

---

🚀 Features

🔎 Text Analysis

Detects phishing, fraud intent, coercion, and social-engineering language.

Models & Methods

• mshenoda/roberta-spam → Scam, phishing, and financial fraud detection
• unitary/toxic-bert → Threats, blackmail, harassment, coercion detection
• Heuristic Risk Engine → Boosts risk when:
  • Urgency language is present
  • Financial pressure is detected
  • Authority impersonation appears
  • Emotional manipulation is used

Output: Risk score + detected scam signals

---

🖼️ Image Analysis

Detects manipulated images and deepfakes using AI + forensic techniques.

Pipeline

• MTCNN → Face detection and alignment
• EfficientNet-B5 → Deepfake artifact detection (GAN textures, blending issues)
• Error Level Analysis (ELA) → Edited region detection via compression inconsistencies
• EasyOCR → Extracts embedded text for scam analysis

Output: Deepfake probability + tampering heatmap

---

🔊 Audio Analysis

Identifies AI-generated or cloned voices using signal forensics.

Techniques

• Spectral Flatness & Centroid → Detect unnatural frequency distributions
• MFCC → Flags missing human micro-variations in speech
• Rule-based Vocal Tract Forensics → Detects synthetic resonance patterns

Output: Human vs AI likelihood + anomaly indicators

---

🎥 Video Analysis

Performs frame-level deepfake detection with temporal consistency checks.

Pipeline

• Keyframe Extraction + EfficientNet-B5 → Artifact detection per frame
• MTCNN Landmark Tracking → Facial stability across frames
• Temporal Flicker Detection → GAN frame instability detection
• Lighting & Shadow Analysis → Physically inconsistent illumination detection

Output: Frame-wise risk timeline + overall deepfake score

---

🧠 Explainable Risk Scoring

• Unified risk score (0–100)
• Modality-wise breakdown (text, image, audio, video)
• Human-readable reason codes for transparency

---

🏗️ System Workflow

1. Input ingestion (text / image / audio / video)
2. Modality-specific analysis pipelines
3. Feature extraction and forensic checks
4. Heuristic risk engine
5. Explainable scoring API
6. Dashboard / Chrome extension output

---

🌐 Use Cases

• Social media scam detection
• Deepfake media verification
• Voice clone fraud prevention
• KYC and identity verification
• Cybercrime and digital forensics

---

🔐 Privacy & Security

• Stateless processing by default (no permanent storage)
• Optional encrypted storage for forensic workflows
• Data minimization and secure inference pipelines

---

⚙️ Tech Stack

• NLP: RoBERTa, Toxic-BERT
• Computer Vision: EfficientNet-B5, MTCNN, ELA
• OCR: EasyOCR
• Audio Forensics: MFCC, spectral analysis
• Backend: Python, FastAPI
• Frontend: React / Chrome Extension
• Deployment: Docker, GPU inference support

---

📊 Example Output

{
  "risk_score": 87,
  "modality_scores": {
    "text": 82,
    "image": 91,
    "audio": 12,
    "video": 0
  },
  "flags": [
    "financial_urgency_detected",
    "face_blending_artifacts",
    "ela_edit_regions"
  ],
  "verdict": "High likelihood scam/deepfake"
}