Posted by Terence Zhang – Developer Relations Engineer, and Adrien Couque – Software program Engineer
AI can improve the person expertise and productiveness of Android apps. For those who’re seeking to construct GenAI options that profit from further information privateness or offline inference, on-device GenAI is an effective selection because it processes prompts immediately in your system with none server calls.
Gemini Nano is probably the most environment friendly mannequin in Google’s Gemini household, and Android’s foundational mannequin for operating on-device GenAI. It is supported by AICore, a system service that works behind the scenes to centralize the mannequin’s runtime, guarantee its protected execution, and defend your privateness. With Gemini Nano, apps can supply extra customized and dependable AI experiences with out sending your information off the system.
On this weblog put up, we’ll present an introductory look into how Gemini Nano and AICore work collectively to ship highly effective on-device AI capabilities whereas prioritizing customers’ privateness and security.
Non-public Compute Core (PCC) compliance
At Google I/O 2021, we launched Non-public Compute Core (PCC), a safe setting designed to maintain your information non-public. At I/O in 2024, we shared that AICore is PCC compliant, that means that it operates beneath strict privateness guidelines. It may possibly solely work together with a restricted set of different system packages which are additionally PCC compliant, and it can not immediately entry the web. Any requests to obtain fashions or different info are routed by way of a separate, open-source companion APK referred to as Non-public Compute Providers.
This framework helps defend your privateness whereas nonetheless permitting apps to learn from the ability of Gemini Nano. Think about a keyboard utility utilizing Gemini Nano for a reply suggestion characteristic. With out PCC, the keyboard would require direct entry to the dialog context. With PCC, the code that has entry to the dialog runs in a safe sandbox and interacts immediately with Gemini Nano to generate strategies on behalf of the keyboard. This enables the keyboard app to learn from Gemini Nano’s capabilities with out immediately accessing or storing delicate dialog information. You will discover out extra about how this works within the PCC Whitepaper.
Defending your privateness by way of information isolation
AICore is constructed to isolate every request to guard your privateness. This prevents apps from accessing information that doesn’t belong to them. Requests are dealt with independently and processed from a single app at a time to mitigate the danger of information being uncovered to different apps.
Moreover, AICore does not retailer any report of the enter information or the ensuing outputs after processing every request. This design, mixed with the truth that Gemini Nano’s inference occurs immediately in your system, helps guarantee your app’s information stays non-public and safe.
Prioritizing Security in Gemini Nano
We’re dedicated to constructing AI responsibly, and that features ensuring Gemini Nano is protected. We have applied a number of layers of safety to restrict dangerous or unintended outcomes:
- Native mannequin security: All Gemini fashions, together with Gemini Nano, are skilled to be safety-aware out of the field. This implies security issues are constructed into the core of the mannequin, not simply added as an afterthought.
- Security conscious fine-tuning: We use a LoRA fine-tuning block to adapt Gemini Nano for the wants of particular apps. Once we practice the LoRA block, we incorporate security information particular to the app’s use case to protect and even improve the mannequin’s security options throughout fine-tuning the place relevant.
- Security filters on enter and output: As a remaining safeguard, each the enter immediate and outcomes generated by the Gemini Nano runtime are evaluated towards our security filters earlier than offering the outcomes to the app. This helps stop unsafe content material from slipping by way of, with none loss in high quality.
These layers of safety work collectively to make sure that Gemini Nano offers a protected and useful expertise for everybody.
Get began
Study extra about Gemini Nano for app growth, and strive it out in your personal app!
Make sure you try the opposite wonderful AI on Android Highlight week content material!
