Local storage exposes models to potential extraction, reverse engineering, and tampering. Attackers can decompile the APK, extract the model file, and analyze its architecture. Mitigation includes encryption, obfuscation, and splitting sensitive processing between device and server.

Implement input validation, use adversarial training during model development, apply input sanitization, monitor for unusual prediction patterns, and set confidence thresholds. Consider using defensive distillation and gradient masking techniques.

On-device inference offers privacy, offline capability, and lower latency but exposes the model. Cloud-based inference provides better security and model updates but requires internet connectivity. Hybrid approaches can balance both needs based on sensitivity and use case.