Gemma Guard is an Android anti-phishing app whose goal is to identify phishing risks before the user taps on suspicious content. It captures the current screen, extracts visible text, and uses a locally running Gemma 4 model to assess suspicious messages, login pages, and payment prompts. It ultimately presents the user with a clear verdict, confidence level, reasoning, and suggested next steps, designed for everyday users and families. The project was built for the Google Gemma 4 Good Hackathon.
Purpose Gemma Guard aims to answer a quick question: is the current screen trustworthy? It targets the kinds of screens that prompt users to tap in an instant, such as fake bank or payment prompts, account lockout or password reset pages, and delivery, refund, or package scams. The app is also suitable for family safety confirmation scenarios, where users can share the result with a trusted contact to get help when a second opinion is needed.
Key capabilities
- Screen capture: captures the current Android screen exactly as the user sees it.
- OCR text extraction: extracts visible text, links, brands, and urgency cues on the screen.
- Local image + text analysis: Gemma 4 evaluates both the screenshot and the OCR text on-device.
- Clear verdict output: displays the verdict, confidence level, reasoning, and the safest next step.
- Runs entirely on-device: screenshot analysis, OCR checks, and reasoning are all completed locally by Gemma 4, with no cloud, and data never leaves the phone, so private messages and sensitive screens stay private.
- Combines images and OCR text: looks at both visual design and text content.
- Confidence scoring and reasoning: users get not just a warning, but understandable context.
- Designed for non-technical users: brief guidance, clear language, and simple next-step suggestions.
- Share results with trusted contacts: involve family when a second opinion is especially needed.
How to use it The app flow is short and easy to understand, suitable for high-pressure moments. First, capture the current screen, then extract visible text via OCR, then Gemma 4 analyzes the screenshot and text on-device, and finally it displays the verdict, confidence level, reasoning, and the safest next step. The entire experience stays visual, readable, and calm. The core phishing check is designed to be local-first to protect privacy and trust.
Use cases
- Fake bank or payment prompts: urgent transfer requests, payment confirmations, and bank warnings.
- Account lockout or password reset screens: pages that urge users to log in quickly before they have time to verify.
- Delivery, refund, or package scams: seemingly harmless everyday messages that make people tap without thinking.
- Family safety confirmation: when a second opinion helps, results can be shared with a trusted contact.
Limitations Gemma Guard's core experience is designed to run on-device, but no information is provided about the specific supported Android versions, device requirements, installation method, or store availability. No fees, pricing, or subscription model are mentioned either. The effectiveness of OCR and Gemma 4 may be affected by the clarity of the screen content, language, and model capabilities, but the material does not provide these specific limitations. The app is not a substitute for professional security software or the user's own judgment, and the material does not provide other limitation statements.
Background Gemma Guard was built by Milana Kerbel and Pavel Kerbel for the Google Gemma 4 Good Hackathon. The project was inspired by Milana's mother, who was scammed by a phishing message last year. Contact: hello@gemmaguard.org.
