Source
How We’re Building Scam Alert on WhatsApp With End-to-End Encryption and Verifiability Guarantees
WhatsApp is introducing Scam Alert, an optional on-device machine learning feature that warns users about potential scam messages from non-contacts.
Why it matters
This allows users to detect fraudulent messages from strangers while keeping their private conversations encrypted and invisible to Meta.
The details
- The tool performs analysis entirely on-device without auto-reporting message content.
- Model weights are public for independent verification by security researchers.
- Users can manually toggle the feature on or off at any time.
Show entities and relationshipsHide entities and relationships
In this article
Key connections
Scam Alert uses Oblivious HTTP (OHTTP)
Scam Alert routes metrics and model download requests through an OHTTP relay that strips IP addresses.
Meta owns Scam Alert
Meta developed Scam Alert as a security feature for WhatsApp.
Scam Alert is related to WhatsApp
Scam Alert is an on-device scam detection feature integrated into WhatsApp.
Cloudflare is a partner of Meta
Cloudflare acts as a third-party signer using Ed25519 keys to sign model manifests for Meta.
Scam Alert uses Machine Learning
Scam Alert runs an on-device machine learning model to classify scam messages.
Scam Alert is related to Cybersecurity
Scam Alert classifies known scam patterns and warns users while preserving end-to-end encryption.
Show 3 more connectionsShow fewer connections
WhatsApp uses End-to-End Encryption
WhatsApp protects personal messages with end-to-end encryption.
Scam Alert uses Differential Privacy
Scam Alert applies differential privacy noise to aggregated telemetry before sending it to servers.
Scam Alert uses Trusted Execution Environment (TEE)
Scam Alert telemetry is processed in confidential virtual machines within Trusted Execution Environments.
Related events
WhatsApp Announces On-Device Scam Alert Feature with Privacy Guarantees
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.