Chipmunk or Pepe? Using Acoustical Analysis to Detect Voice-Channel Fraud
summary
Detects voice‑channel takeover by combining 1380 audio features (non‑voiced noise spectra, voiced vocal‑tract models, signaling/ANI velocity) into an audio fingerprint or phone print that flags ANI spoofing, VoIP gateway abuse and voice distortion. Claims: 731M calls analyzed, 942k fraud events, 43k attackers, 210% rise since 2013 and ~1/638 calls fraudulent in 2017. System runs streaming feature extraction and ML in real time (5–7s of audio, ~300ms scoring), shares detections across customers via a consortium.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:24Three ways you interact with the business
- 2:08The most popular channel
- 6:26Apple Pay
- 7:17First Call
- 9:14The Problem
- 10:24The Data
- 11:36Fraud Rates
- 20:14Chipmunk
- 23:21Techniques Used
- 28:53Examples
- 30:27VoiceOverIP Systems
- 31:11Vocal Tract
- 34:58Engineering
- 39:17Blackboard Architecture
- 40:04Rock Tumbler
related talks
53mCall Me: Gathering Threat Intelligence on Telephony Scams to Detect Fraud
54mLifecycle of a Phone Fraudster: Exposing Fraud Activity
45mVoice Security - Get Up to Speed on This Attack Vector
39mSynthetic Trust: Exploiting Biases at Scale
33mYour voice confirms my identity
20mSecurity in the Age of AI: Response and Strategy
06mVoice In Ear: Spoofing-Resistant and Passphrase-Independent Body Sound Authentication
39m Put Words In My Mouth