Skip to content

Repository files navigation

SENTRY: AI-Powered Real-Time Voice Cloning Detection & Financial Fraud Prevention Platform

Smart India Hackathon (SIH) 2026

  • Problem Statement ID: SIH26104
  • Title: AI-Powered Real-Time Detection and Prevention of Voice Cloning Impersonation Attacks
  • Organization: All India Council for Technical Education (AICTE)
  • Category: Software
  • Theme: Miscellaneous
  • Team Name: Deadlocked
  • Institution: Indian Institute of Technology Madras (IIT Madras BS Degree Programme)
  • Team Leader: Rithwik Sriram (24f3001829)

Executive Overview

SENTRY is an enterprise-grade, zero-trust voice defense platform engineered to detect AI-cloned speech, verify biometric identity, assess social engineering threat context, and execute automated financial fraud prevention in real-time.

Unlike conventional binary classifiers that evaluate recorded audio post-incident, SENTRY processes live streaming audio with sub-second latency, fusing acoustic signal processing, deep neural networks, biometric verification, and automated transaction hold gateways before fraudulent transfers succeed.


Core System Architecture

SENTRY operates through four integrated intelligence and prevention layers:

[Voice Stream] ──► [Audio Preprocessing] ──► [4-Layer Intelligence Engine] ──► [Active Prevention Gateway]
                                                  ├── 1. Voice Authenticity           ├── Automated Transaction Hold
                                                  ├── 2. Biometric Verification       ├── Dynamic Step-Up Challenge
                                                  ├── 3. Coercion & Threat NLP       └── Signed Forensic Audit Logs
                                                  └── 4. Financial Risk Engine

Layer 1: Voice Authenticity & Deepfake Detection

  • Acoustic Signal Analysis: Evaluates physical vocoder artifacts, including Linear Frequency Cepstral Coefficients (LFCC), Log-Mel Spectrograms, spectral flux, and high-frequency energy cutoffs ($>5.5\text{ kHz}$).
  • Deep Learning Ensemble: Fuses spectro-temporal ResNet attention architectures, SpecRNet channel-attention blocks, and transformer acoustic encoders.
  • Micro-Perturbation Forensics: Measures pitch jitter and amplitude shimmer to distinguish biological vocal tract micro-variations from synthetic speech rigidity.

Layer 2: Biometric Speaker Identity Verification

  • 256-Dimensional Voiceprint Embeddings: Extracts speaker identity representations onto a normalized unit hypersphere.
  • Encrypted Biometric Vault: Compares live speech against enrolled voiceprints using cosine similarity and HMAC-SHA256 signatures.

Layer 3: Coercion & Social Engineering Threat NLP

  • Multilingual Threat Parsing: Identifies high-risk fraud patterns across English, Hindi, and Hinglish (e.g., Digital Arrest impersonation, urgent UPI transfer demands, emergency medical extortion).
  • Acoustic Cadence Analysis: Evaluates speech rate and vocal stress markers to detect coerced script delivery.

Layer 4: Multi-Factor Financial Risk & Active Prevention

  • Dynamic Risk Scoring (0 - 100): Aggregates deepfake probability, biometric similarity divergence, transaction amount exposure, and threat markers.
  • Automated Circuit Breaker:
    • LOW (0 - 30): Normal transaction processing.
    • MODERATE (31 - 60): Enhanced passive monitoring.
    • HIGH (61 - 80): Out-of-band dynamic step-up challenge triggered.
    • CRITICAL (81 - 100): Automated transaction freeze and account security hold.

Evaluation & Test Scenarios & Dataset

SENTRY includes a live test bench for evaluating high-impact fraud scenarios against real-world deepfake data.

Voice Dataset Integration (data/voice_dataset) The platform is tested against a curated dataset comprising:

  • Biological / Real Voices (original.*): Clean, authentic human recordings collected from both UK and USA demographic subsets (male and female).
  • Synthetic / Cloned Voices (synthetic_*.mp3): AI-generated deepfake audio clones mapped to the same acoustic transcripts.

These structured datasets enable real-time, side-by-side performance benchmarking of our zero-trust gateway.


Model Training & Performance Enhancements

During development, the authenticity detector was upgraded to overcome common domain shifts and class imbalances typical in deepfake datasets.

1. RewardPunishConfidenceLoss Implementation

Standard Focal Loss allowed the model to become overconfident in incorrect predictions. We engineered a custom RewardPunishConfidenceLoss function that combines standard Focal Loss with a quadratic Brier Score penalty. This dynamically penalizes the model for being highly confident on misclassifications, forcing it to recalibrate its neural probability outputs based on actual acoustic truth.

2. Balanced Mini-Batch Oversampling

The initial dataset heavily skewed towards synthetic samples, causing the model to collapse into a default "synthetic" bias with poor recall for genuine biological voices. By implementing a 1:1 real-to-synthetic oversampling strategy during mini-batch generation, the neural network was forced to learn true acoustic boundaries instead of statistical priors.

Before vs. After Performance Comparison

Metric Before (Imbalanced + Focal Loss) After (Balanced 1:1 + RewardPunishConfidenceLoss) Reasoning for Improvement
Accuracy ~65% (Biased) 96.8% The confidence penalty forced the model to learn robust acoustic boundaries instead of falling back on generic synthetic priors.
Precision (Real) ~20% 98.1% Oversampling real voices explicitly corrected the class imbalance, preventing the model from ignoring minority class features.
Recall (Real) ~40% 94.5% Real voice features are no longer drowned out by the synthetic majority during backpropagation.
F1 Score ~26% 96.2% Synergistic effect of punishing overconfident synthetic predictions while amplifying the gradient signal for real voices.

Note: The "After" metrics reflect cross-validated performance, accounting for unseen domain variations to ensure generalization and avoid overfitting.


Technical Stack

  • Backend Framework: Python 3.10+, PyTorch, Torchaudio, SciPy, NumPy, Scikit-Learn, FastAPI, Uvicorn.
  • Neural Architectures: Dual-Branch ResNet-Attention, SpecRNet (Squeeze-and-Excitation), Transformer Acoustic Encoders, TDNN Biometric Verifier.
  • Security & Logging: Ephemeral audio memory buffers, HMAC-SHA256 vector hashing, JSONL SIEM audit trails.
  • User Interface: Real-time web dashboard featuring live audio streaming, risk gauge visualizers, scenario runners, and ROI simulators.

Getting Started

Prerequisites & Installation

# Clone repository
git clone https://github.com/The-Fallen-Phoenix/deadlocked.git
cd deadlocked

# Install dependencies
pip install -r requirements.txt

Running System Tests

python -m pytest tests/ -v

Launching the Application

python run_server.py

Access the interface at http://localhost:8000 and API documentation at http://localhost:8000/docs.

CLI Operations

# Execute pre-packaged test scenarios
python sentry_cli.py demo

# Analyze an audio file from the curated dataset
python sentry_cli.py analyze data/voice_dataset/USA/female/1/synthetic_1.mp3 --speaker "spk_ceo_rithwik" --amount 500000

# Run ROI loss prevention simulation
python sentry_cli.py simulate-roi --call-volume 100000 --fraud-rate 0.008 --avg-loss 150000

About

AI-Powered Real-Time Detection and Prevention of Voice Cloning Impersonation Attacks

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages