Runtime security boundaries for mobile AI applications
Yudun technical guide · Engineering principles and practice · English
A mobile AI application often combines models, inference runtimes, SDKs, interfaces, and business decisions. Map where each asset is created, transferred, used, and updated before deciding what the client should protect.
Treat the model as one asset among many
Inventory prompts, preprocessing, postprocessing, interface credentials, caches, logs, and user inputs and outputs alongside the model.
Each asset has different confidentiality, integrity, and availability needs.
- Models and configuration
- SDKs and runtimes
- Interfaces and credentials
- Inputs, outputs, and logs
Split client and server responsibilities
The client can support low-latency and offline inference. Long-lived secrets, account authorization, and high-risk business decisions are safer under server control.
Client signals are evidence, not absolute trust.
Secure the update path
Models, runtimes, and application versions require defined compatibility. Verify update source, integrity, rollback, and rollout range.
The system should be able to stop a defective model or policy instead of keeping it indefinitely.
- Version pairing
- Update source
- Integrity verification
- Failure rollback
Publish within the evidence boundary
Without a current sample and runtime evidence, publish threat models, design principles, and acceptance checklists, not claims that extraction or memory modification was blocked.
Verify the boundary with a real release candidate
Provide the application stack, critical paths, and target system range. Accounts, applications, and project details are handled by the central Yudun platform.