Google DeepMind brings on-device Gemma to fight conversational scams

Scammers don't just blast out malicious links anymore; they play the long game. Using sophisticated social engineering, they chat up victims for days before ever asking for cash. In Singapore, these conversational scams account for roughly 80% of the S$4 billion lost to fraud since 2019.

On October 14 at Singapore International Cyber Week (SICW 2026), Google DeepMind and Singapore's Government Technology Agency (GovTech Singapore) showcased a fix: an on-device, fine-tuned Gemma model designed to detect manipulation in real time without snooping on private messages.

A two-stage pipeline for on-device detection

Traditional threat detection relies heavily on cloud servers scanning text. That model falls apart with end-to-end encrypted messaging, where cloud scanning breaks privacy promises.

DeepMind solved this by pushing the intelligence directly onto the smartphone. To keep the model from draining the battery, the team implemented a two-stage pipeline.

Stage one uses an ultra-lightweight classifier that monitors inbound messages with minimal power draw, screening for early red flags.

Stage two triggers only when something looks fishy. The fine-tuned Gemma model kicks in to analyze a broader snapshot of the multi-turn conversation, assessing manipulative intent to confirm a threat or discard false alarms.

If a scam is detected, the device pops an instant alert advising the user to pause before sharing sensitive credentials or sending funds. Everything runs strictly on-device with explicit user consent; neither Google nor the Singapore government ever sees the chats.

Fine-tuning on local tactics inside a sovereign sandbox

Generic foundation models usually miss local nuances and slang, which makes detecting regional fraud tough.

To bridge that gap, the team fine-tuned Gemma on localized scam scripts, including impersonations of local public agencies, regional banks, and trending neighborhood schemes. This allows the model to separate harmless small talk from actual financial grooming.

Governance was kept tight. Fine-tuning took place entirely inside a sovereign sandbox on GovTech’s managed GCP environment.

No personal citizen data was used; training relied strictly on anonymized scam reports, threat intelligence feeds, and synthetic dialogues tested against diverse local corpora.

Why it matters

We think this project offers an elegant blueprint for how platforms can balance consumer protection with end-to-end encryption. For years, cybersecurity debates treated privacy and safety as a zero-sum game. Either messages were private and platforms were blind to fraud, or messages were scanned in the cloud at the expense of user privacy. Moving fine-tuned small language models to the edge proves you can catch conversational scams without breaking encryption.

For everyday users, having a local watchdog that understands regional manipulation tactics is massive. Unlike static spam filters that check phone numbers, this actively analyzes the context of a conversation right on your phone, giving you a second opinion before you wire money to an imposter.

The catch? False alarms and adversarial adaptation. If the model interrupts genuine conversations too often, users will simply toggle it off. Moreover, running a static model locally means clever scammers can download it, reverse-engineer its trigger thresholds, and tweak their scripts to evade detection. To stay ahead, edge-deployed models will need constant, lightweight behavioral updates.