Meta Tests Optional WhatsApp Scam Alert with On-Device AI

Meta has commenced the trial of an optional Scam Alert feature for WhatsApp users. This protection, powered by a machine learning model operating directly on devices, follows a recent campaign of WhatsApp account access scams not requiring user passwords. The new system aims to provide warnings about potential fraudulent messages received from unknown contacts.

Credit: UNSPLASH

According to a Forbes report, Meta's engineering team confirmed the feature complements existing end to end encryption. Scam Alert is currently in a limited beta programme. Meta collaborates with the WhatsApp bug bounty community to rigorously test the system. Details regarding wider availability or how individuals might participate in the beta phase have not been released at this time.

Big Tech organisations are actively seeking to fortify communication platforms against ongoing threats. Google recently unveiled a fake call detection system that verifies call legitimacy when both parties use Phone by Google. And it already offers an artificial intelligence powered scam detection feature for Android users. This Google system provides warnings and analyses calls and messages in real time, with all processing occurring on the device to maintain conversation privacy.

Meta's approach with Scam Alert mirrors this on device processing method. The engineering team specified that no message content leaves the user's device for classification or automatic reporting to WhatsApp, Meta, or any other party. This data privacy is maintained because the machine learning models are designed to be small enough to run entirely on any device, requiring no server side components.

The Scam Alert functionality remains entirely optional for users. Once activated, the machine learning model downloads to the device. So it then runs inferences to categorise whether incoming messages from non contacts align with known scam patterns. These patterns have been established through training the artificial intelligence on reported scam instances from WhatsApp users. Meta stated that this probabilistic classification relies on conversational structure and linguistic signals.

If a potential scam is identified, Scam Alert issues a warning to the user. This notification is not visible to the message sender. Users presented with an alert can choose to block the sender, report the chat, or proceed with the conversation. If a user determines a warning was wrongly flagged, they can mark the chat as trusted. This action removes the current warning and prevents Scam Alert from flagging that specific chat again in future.

Scam Alert represents another layer within the established Swiss Cheese security model. This defence in depth strategy involves layering multiple security measures. The individual "holes" in each layer are randomly positioned and sized. But by adding more layers, it becomes progressively more difficult for a threat to penetrate the entire security stack.

Other security layers already provided for WhatsApp users include a recent update to security tools. This update can inform users of potential device linking code attacks. Furthermore, the optional Strict Account Settings configuration automatically silences calls from unknown numbers, blocks attachments from unrecognised users, and prevents the display of link previews.

• Meta has introduced an optional on device Scam Alert feature for WhatsApp users.

• The system uses a machine learning model to identify potential scams from non contacts.

• All message content processing occurs locally on the user's device, ensuring privacy.

• Users can block, report, or continue chats flagged by the system, and can mark chats as trusted.

• The feature adds another layer to WhatsApp's existing security measures.

Source: Forbes