🛡️ Real-Time Audio Fraud Detection • AI vs AI Defense • Conversation Intelligence

---

🛡️ FemtoGuard – Team SheStorm

Real-Time Audio Fraud Detection for Scam Prevention
Conversation Intelligence for the AI vs AI Era (2026)

---

🌐 Live Demo

🔗 App:
https://shestorm-ai-fraud-defender-73291669658.us-west1.run.app

📂 Demo + PPT:
https://drive.google.com/drive/folders/1y_DknpPaxDXdqYZMj07zlWCOonYMsOap

---

👥 Team SheStorm

Name	Role
Yamini	Frontend & UX
Ishani Gupta	Backend & APIs
Madhu Tiwari	AI / ML
Khushi Verma	Research & Testing

---

🚨 Problem

Voice fraud has evolved into AI-driven psychological manipulation:

• 🎭 Voice cloning in seconds
• 🤖 AI-driven scam conversations
• 📞 Caller ID spoofing
• 🧠 Emotional exploitation

❌ Traditional systems ask:

“Is the voice fake?”

✅ We ask:

“Is the intent malicious?”

---

💡 Our Approach

FemtoGuard detects fraud in real-time by analyzing:

• 🧠 Intent
• 🎭 Behavior
• 💬 Conversation patterns

📌 We don’t detect the caller — we detect the conversation itself.

---

🧠 Core Idea

🔄 Identity → Intent Shift

Traditional Systems	FemtoGuard
Who is calling?	Why are they calling?
Is voice real?	What are they asking?
Known number?	How are they manipulating?

---

🔍 Detection Engine

1️⃣ Intent Detection

• Authority phrases (bank, officer)
• Urgency cues (immediately, now)
• Financial triggers (OTP, PIN)
• Isolation tactics

---

2️⃣ Behavioral Analysis

• Scripted speech patterns
• Repetition loops
• Interruptions
• Dominant tone

---

3️⃣ Emotional Manipulation

• Fear induction
• Pressure tactics
• Aggression mismatch

📌 “Bank agent threatening user” = 🚨 High Risk

---

🧠 AI Pipeline

flowchart TD
    A["🎙 Audio Stream"] --> B["📊 Feature Extraction"]
    B --> C["🧠 Speech-to-Text"]
    
    C --> D["📌 Intent Detection"]
    C --> E["🎭 Behavior Analysis"]
    C --> F["💥 Emotion Detection"]
    
    D --> G["⚠️ Risk Engine"]
    E --> G
    F --> G
    
    G --> H["🚨 Real-Time Alerts"]

Loading

---

⚙️ System Architecture

Audio Input
   ↓
Feature Extraction
   ↓
Transcription
   ↓
Intent + Behavior + Emotion Analysis
   ↓
Risk Scoring
   ↓
User Alert System

---

📊 Before vs After Fraud Detection

🎯 Scenario: Scam Call Attempt

---

❌ Before (Traditional Systems)

Call Transcript:

"Hello ma'am, I am calling from your bank. Your account will be blocked immediately. Please share your OTP to verify."

System Response:

• Caller ID: Unknown ❓
• Voice: Human-like ✅
• Blacklist match: ❌

📌 Result: No alert
🚨 User Outcome: High scam risk

---

✅ After (FemtoGuard)

Signal Type	Detection
Authority Claim	Bank detected
Urgency Cue	Immediate
Financial Trigger	OTP
Tone Analysis	Aggressive

Risk Score: 92% (HIGH RISK)

🚨 Alert:

⚠️ "Potential scam detected. Do NOT share sensitive information."

---

💡 Impact

Aspect	Before	FemtoGuard
Detection	Caller-based	Intent-based
Speed	Slow	Real-time
Accuracy	Low	High
Protection	❌	✅

---

🛠 Tech Stack

🧠 AI / ML

• Speech features (MFCC, spectrogram)
• NLP / LLM models
• Real-time inference

⚙️ Backend

• FastAPI
• WebSockets
• REST APIs

🎨 Frontend

• Live dashboard
• Risk meter
• Alerts

🗄 Database

• PostgreSQL / SQLite

---

✨ Key Features

• ⚡ Real-time fraud detection
• 🔐 No prior enrollment
• 🧠 AI + human scam detection
• 🌍 Works on first call
• 🔊 Noise tolerant

---

🧪 Dataset

• Synthetic scam conversations
• Multi-language support
• Emotional variations

---

🚀 Future Scope

• 📱 Mobile integration
• 🌍 Multilingual support
• 📡 Telecom deployment
• 🧠 Deep learning upgrades

---

🧑‍💻 Run Locally

npm install

Add API key in .env.local

npm run dev

---

🏁 Conclusion

Voice fraud is not an audio problem.
It is a human manipulation problem.

🛡️ FemtoGuard acts as a Real-Time Conversation Firewall
— stopping fraud before damage happens.

---