Singapore is battling a surge in fraudulent activity. The city-state saw a record high of 15,756 scams reported last year, with more than S$201 million (US$151.3 million) lost.
The country has launched an AI-powered app that can identify, filter, and block scam messages with 95 per cent accuracy. Known as ScamShield, the app was built by the Open Government Products (OGP) team, a unit in the Government Technology Agency, and the National Crime Prevention Council.
GovInsider spoke to Lennard Lim, ScamShield’s Product Manager, to find out how the app tackles scams while protecting privacy.
How it works
The app filters scam messages into a junk folder using AI. ScamShield can also block calls from numbers reported by users, or those on a list maintained by the Singapore Police Force.
The model is trained to recognise common words used in fraudulent texts such as ‘loans’, ‘gambling’, and ‘repayments’, Lim says. It then determines whether a text is fraudulent by looking at how these words are used in combination with one another.
The ScamShield team trained the model with scam messages they had gathered from volunteers and publicly available sources, Lim says. They then generated different variations of scam messages to further enhance the model’s detection capabilities.
To make the app user-friendly, the ScamShield team decided to make it function with “minimal user intervention”. Users only have to grant the app permission to block SMS and calls, and the app will work “seamlessly in the background”.
The team decided to block and move scam messages to a junk folder instead of exposing users to the message. This is useful for the elderly, Lim explains. Scammers are getting more sophisticated, and there can be little difference between a fraudulent and real message.
Scammers “change their attacks quickly” and it’s crucial for the model to be able to adapt. The team retrains the model regularly and updates its repository of scam templates when new ones are detected, he adds.
OGP also works closely with police, banks, and other agencies to be immediately notified of new scam messages and calls. Users are encouraged to report scam messages and calls that the model failed to detect.
722,865 SMSes have been reported, and more than 5,537 phone numbers have been blocked on the ScamShield app since its launch six months ago, according to the Singapore Police Force and National Crime Prevention Council.
Ensuring user privacy
The app was built with users’ privacy in mind, Lim says. It has two features to ensure that.
First, Apple’s Message Filtering framework ensures only messages from unknown senders are scanned by the app. That prevents it from mistakenly screening any private messages from friends and family, he says.
Second, the model analyses messages on the users’ device and not on OGP’s servers. Only scam messages are forwarded to the ScamShield team, and iOS handles this forwarding.
This ensures only the sender and message content are shared. ScamShield does not and cannot know which devices forwarded the message, Lim explains.
The OGP team considered some common approaches to tackling scams when building the app. The US is looking to implement a technology that verifies caller ID, but this requires infrastructure changes that take a lot of effort and time, he says.
Other apps such as TrueCaller come with a subscription cost and require the calls and texts to pass through their servers. That may compromise user privacy, Lim adds.
By running the algorithms on users’ devices, the team “made a conscious decision” to respect users’ privacy. This prevents them from knowing how users are interacting with the app, but Lim shares that they still collect user feedback via email and App Store reviews.
The app has come a long way from when it was first conceived during OGP’s Hackathon last January.
The team is now looking at shortening the time between detecting new scam types and ensuring they are filtered by the app, Lim says. They also plan to work with the police to alert users via in-app notifications whenever a new scam outside of the app’s domain is detected, he adds.
OGP hopes to launch an Android version by the end of the year to cater to a wider range of users. It had planned to extend ScamShield to WhatsApp and Telegram, but that has been difficult as the platforms are encrypted, Lim shares.
The team, however, does plan to create an additional channel for users to report WhatsApp and Telegram scams to the police and National Crime Prevention Council, he adds.
ScamShield’s model is not open sourced at the moment, but the team intends to share the code in the “near future”.
“We believe that for every message blocked, it reduces the chance that someone’s savings are cheated.” The ScamShield app will play a crucial role in helping Singapore tackle its rise in scams.
Images from the Open Government Products team