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

12 Aug 2026

Problem

WhatsApp messages are end-to-end encrypted, so the server cannot run a scam classifier on them. The team still has to know whether the filter catches real scams, how often it raises false alarms, and when to update the model as scam tactics change. Outside researchers and users must be able to check that no message content leaves the phone and that nobody can ship a special model to one person.

Example from the article: Scam Alert warns a user about a likely scam from a non-contact, and the user can block, report, continue, or mark the chat as trusted. Only anonymous, noisy counts of warnings and actions reach WhatsApp.

Solution in brief

Classification runs on the device. Quality is measured from two counts (warnings shown, user actions), aggregated locally and sent through an anonymizing relay into a hardware-isolated environment that releases only differentially private totals. Every model version is signed by a third party and recorded on a public append-only ledger before download, checked by the client.

Questions this case answers
  • How to measure the quality of an on-device classifier when raw inputs and per-user data must never reach the server
  • How to roll out and experiment with model versions so that no one can target an individual user and every version is publicly auditable
  • How to warn users about scam messages in an end-to-end encrypted chat without sending message content to the server?
  • How to measure precision and false positives of an on-device classifier using only privacy-preserving aggregate counts?

Tools and methods

  • On-device text classifier — flags messages from non-contacts that match scam patterns
  • Confidential federated analytics — on-device counts, TEE aggregation, k-anonymity, noise
  • Differential privacy — noise so one person's data barely changes a released number
  • OHTTP relay and anonymous credentials — hide the client IP and identity
  • Transparency ledger with third-party signing — tamper-evident record of model versions
Continue reading — it's free

Google or a one-time email code. No password.