Google DeepMind + GovTech Singapore Ship On-Device Gemma for Scam Detection

Google DeepMind and Singapore’s Government Technology Agency (GovTech Singapore) shipped a fine-tuned, on-device Gemma model for conversational scam detection on Oct 14, 2026 — presented at Singapore International Cyber Week (SICW 2026) as part of their National AI Partnership. The architecture is privacy-first by construction: the entire inference runs on the user’s device with zero cloud data egress, neither the Singapore Government nor Google can read user conversations through deployment, and the fine-tuning happened inside GovTech’s managed GCP “sovereign sandbox.” This is the rare public-sector AI deployment that doesn’t trade privacy for protection. (DeepMind blog, Unite.AI)

The deployment pattern matters more than the model itself. Most anti-scam tools today route conversations to a server-side scanner — which is exactly what kills end-to-end encryption and pushes governments and providers into a standoff with privacy advocates. DeepMind’s design dodges that entirely by moving the classifier to the edge.

The problem: $4 billion in losses, 80% from social engineering

Singapore’s reported crimes are over half scams, with more than S$4 billion in cumulative losses since 2019. Scam losses in the first half of 2026 alone hit S$410.6 million — down from the same window in 2025, but still averaging more than S$68 million per month. The structural shift is the harder problem: simple phishing has been replaced by complex conversational social engineering, where attackers build rapport over multiple messages before asking for money or credentials. Up to 80% of all scam losses now come from this category. (Straits Times H1 2026, Straits Times scam tracker, DeepMind blog)

The “8 in 10 victims voluntarily handed money to scammers” stat from the Straits Times is the key behavioral signal: the attack is now persuasion over multiple turns, not credential phishing. Detecting it requires reading the shape of the conversation — escalating urgency, roleplay of a bank officer, requests that break normal procedures — not just URLs.

The model: lightweight Gemma, fine-tuned on local scam corpora

DeepMind took Gemma (the open-weight family) and fine-tuned it on local scam conversations specific to Singapore’s threat landscape: impersonations of public institutions, banks, and local figures, plus trending regional schemes (job scams, investment fraud, fake officials). The fine-tuning data was derived from anonymised scam reports and intelligence feeds — no private citizen data, no PII. (DeepMind blog, Unite.AI)

The published blog doesn’t specify which Gemma variant (size, modality) or the parameter count. It also doesn’t disclose the fine-tuning dataset size, evaluation metrics, or false-positive rate against benign social chatter. Those numbers are pending verification.

The architecture: two-stage pipeline, on-device

The on-device constraint is the hard part. Running a multi-billion-parameter Gemma model on every inbound message would burn battery and latency. So DeepMind’s design splits the work into two stages:

Stage	Model	Job	Trigger
1	Lightweight classifier	Screens inbound messages for potential threat patterns using minimal power	Always-on per message
2	Full fine-tuned Gemma	Analyzes a larger conversation snapshot to confirm threat vs. discard false alarm	Stage 1 flags

Both stages run on-device, with no cloud round-trip. When Stage 2 confirms a threat, the app shows a real-time warning to the user, advising them to pause before sharing funds or credentials. (DeepMind blog)

This is the canonical “small filter, big confirmer” pattern for edge inference — Stage 1’s job is to keep Stage 2 from running constantly. The DeepMind blog doesn’t publish the Stage 1 model’s specs or the Stage 1 → Stage 2 escalation rate, both of which are the interesting engineering numbers for anyone trying to replicate the design.

Privacy architecture: the part that matters

Four structural choices make this deployment genuinely privacy-preserving, not a privacy-washing press release:

On-device only: As the model runs entirely on-device, neither the Singapore Government nor Google will be able to access users’ conversations through its deployment. No cloud API call carries conversation content.
Sovereign sandbox: Fine-tuning took place inside GovTech Singapore’s managed and maintained GCP environment, keeping training data isolated.
No citizen data: Training strictly excluded private or personally identifiable information, using only anonymised datasets.
Stress-testing surface area: The model was tested against synthetic dialogues and local corpora to cover diverse scam categories and evolving impersonation tactics.

The deployment will only access on-device conversations with user consent, in accordance with the application’s terms of use. (DeepMind blog)

The DeepMind blog explicitly frames this as a model for messaging platforms balancing user protection with end-to-end encryption. That’s the strategic point — Signal, WhatsApp, iMessage, and other E2EE platforms have all been under pressure to scan messages server-side for child safety and scam content. An on-device classifier removes the need to compromise encryption.

What shipped and what didn’t

What shipped: A fine-tuned Gemma variant designed to run on-device in two stages. A governance framework (sovereign sandbox, no-PII training, stress-testing). Presented at SICW 2026 on Oct 14, 2026.
What didn’t ship (or wasn’t disclosed): A specific app or platform integration. The model is being phased in partnership with the Singapore Government — no timeline for general availability was published in the DeepMind blog post. No benchmark numbers (latency, accuracy, false-positive rate, battery cost). No disclosure of which Gemma base size or fine-tuning dataset size. All four are pending verification.
What’s adjacent: The Singapore Government will need to navigate the same E2EE debate that the EU is hashing out under the CSA Regulation, and the UK under the Online Safety Act. A working on-device scam classifier could be the template that lets E2EE platforms say “we can do scam detection without breaking encryption.”

Why this deployment pattern is the part worth watching

Most production AI deployments follow the same shape: cloud API, central inference, content goes to the provider. The privacy guarantees are then a function of the provider’s policies and contracts, which can change. On-device inference inverts that: the privacy guarantee is a function of the architecture, not the policy. If the model runs entirely on the device, there is no cloud-side conversation content to leak, subpoena, or change terms on.

That’s the pattern that’s actually scarce. Most “on-device AI” in 2026 marketing is small-model routing or summarization, not full classifiers. DeepMind and GovTech Singapore shipping a fine-tuned scam-detection model that runs entirely on-device — with explicit “neither party can read user conversations” framing — is the structural milestone. The story to watch in 2027 is whether this pattern gets adopted by WhatsApp, Signal, iMessage, or the EU’s CSA-mandated scanning regime.

Google DeepMind + GovTech Singapore Ship On-Device Gemma for Scam Detection — Zero Cloud Egress