# How We’re Building Scam Alert on WhatsApp With End-to-End Encryption and Verifiability Guarantees

[Meta](https://yomu.fyi/company/meta) · Aug 12, 2026

## Summary

WhatsApp is introducing Scam Alert, an optional feature designed to detect scam messages from non-contacts while preserving end-to-end encryption guarantees. The system executes text classification entirely on-device using a machine learning model trained on conversational patterns from user reports, ensuring raw message content never leaves the client. To evaluate feature accuracy without compromising privacy, the client transmits minimal pre-aggregated telemetry—consisting only of warning and action counts—over Oblivious HTTP relays to Trusted Execution Environments. These confidential virtual machines securely aggregate metrics and apply differential privacy noise before exposing population-level statistics to servers. System integrity and non-targetability are maintained by requiring clients to attest backend binaries against public transparency ledgers before transmitting any telemetry.

## Takeaways

- Scam Alert executes probabilistic text classification strictly on-device for incoming messages from non-contacts, preventing message content from being automatically reported to Meta or WhatsApp.
- Performance telemetry is routed through Oblivious HTTP relays and anonymous credentials into Trusted Execution Environments, which apply differential privacy noise and k-anonymity thresholds before aggregation.
- Targeted model delivery is prevented by publishing all model variants, privacy parameters, and confidential computing binaries to public transparency ledgers verified by on-device attestation.

**Tags:** [Architecture](https://yomu.fyi/topic/architecture), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Privacy](https://yomu.fyi/topic/privacy)

[Read original post](https://engineering.fb.com/2026/08/12/security/how-were-building-scam-alert-whatsapp)