Deep
Learning
Real-Time Fraud Detection in Phone Calls Using
Deep Learning: A Smart Approach to Security
Group No.: B-12 Batch: B2 Class: TY-AI-B
P A G E
Guide: Prof. Shweta Kambare
0 1
studio shodwe
Group Members: 19 Ankita Kakade 12210626
53 Nujaim Maindargi 12210423
55 Manesh Mahale 12210303
57 Mayur Agarwal 12211654
72 Dhanshree Shinde 12320188
P A G E
0 9
studio shodwe
Introduction
Fraudulent phone calls have increased significantly,
leading to financial losses and security threats.
Conventional fraud detection methods are
inefficient in real-time scenarios.
Develop an AI-based real-time fraud detection
system that processes live phone calls, detects
P A G E
fraudulent activity using Deep Learning, and alerts
users instantly.
0 2
studio shodwe
Problem
statement
Traditional fraud detection methods are slow, unreliable,
and rule-based, making them ineffective against evolving
scam tactics. Additionally, existing solutions lack real-
time intervention, failing to prevent fraud as it happens.
P A G E
0 3
studio shodwe
Objectives
Real-Time Fraud Detection – Identify fraudulent phone calls as
they happen.
Automatic Speech Recognition (ASR) – Convert live call audio to
text.
Deep Learning-Based Classification – Use BiLSTM + Attention for
accurate fraud detection.
Instant User Alerts – Notify users with fraud warnings.
Optional Call Termination – Allow users to enable auto-
disconnect for fraudulent calls.
By achieving these objectives, the system enhances security,
accuracy, and user protection against fraud.
studio shodwe
Methodology
The proposed system follows a structured pipeline:
📌 Step 1: Audio Processing
Convert real-time phone call audio into text using OpenAI Whisper ASR.
📌 Step 2: Text Preprocessing
Lowercasing, stopword removal, lemmatization, and special character
removal.
Convert text into BERT embeddings (768-dimensional vectors).
📌 Step 3: Fraud Classification using Deep Learning
Use BiLSTM + Attention model to detect fraudulent conversations.
📌 Step 4: Real-Time Alerts & Call Termination
P A G E
If fraud is detected, trigger an instant warning pop-up.
Optionally, terminate the call if the feature is enabled.
0 7
studio shodwe
Dataset & Preprocessing
Dataset:
📌 10,000 phone call transcripts labeled as fraud or non-fraud.
Preprocessing Steps:
🔹 Convert text to lowercase.
🔹 Remove stopwords, special characters, and numbers.
🔹 Apply lemmatization for text normalization.
Feature Extraction:
🔹 Use BERT embeddings (768-dimensional vectors).
🔹
P A G E
Mean-pool embeddings for fixed-length representation.
0 6
studio shodwe
Model Architecture
📌 Deep Learning Model: BiLSTM + Attention
🔹 Embedding Layer: Pre-trained BERT embeddings (768 dimensions).
🔹 BiLSTM Layer: Captures long-term dependencies in text.
🔹 Attention Mechanism: Focuses on crucial words in the transcript.
🔹 Dense Layer (Output): Sigmoid activation for fraud probability.
P A G E
1 0
studio shodwe
Training Pipeline
✅ Data Split:
80% Training
10% Validation
10% Testing
✅ Loss Function: Binary Cross-Entropy (for fraud classification).
✅ Optimizer: Adam (learning rate = 0.0001).
✅ Evaluation Metrics: Accuracy, Precision, Recall, F1-Score (for handling
class imbalance).
P A G E
0 5
studio shodwe
Model Deployment for
Real-Time Detection
📌 Audio Processing: Convert live call audio to text using
OpenAI Whisper.
📌 Fraud Detection: Use trained BiLSTM + Attention model for
classification.
📌 User Alerts:
⚠ Pop-up alert: “Fraud Call Detected!”
🔴 Automatic call termination (if enabled in settings).
P A G E
1 1
studio shodwe
Conclusion
P A G E
1 4
studio shodwe
Thank
You!
P A G E
1 5