SEAL: Secure and Efficient Adaptive Layering for On-Device Language Models with TEE
Overview
Deploying a small language model on an edge device raises a practical tension between model protection and inference cost. Running in the rich execution environment leaves model parameters exposed to theft, while placing the entire model inside a trusted execution environment can impose substantial latency and memory overhead. SEAL studies how to distribute model layers between these environments under device constraints.
Confidential Layer Analysis evaluates the confidentiality importance of individual layers. A Layer Importance-Guided Adaptive Partition algorithm then uses these assessments to select sensitive layers whose protection fits the available resources. The selected layers execute inside the trusted environment, while other computation remains in the rich execution environment. This makes the partition depend on both the value of protecting a layer and the cost of doing so.
The evaluation considers reconstruction-based model theft together with latency, memory use, and energy consumption. By examining selective protection alongside unprotected and fully protected execution, the work characterizes the security–efficiency trade-off of layer placement. Its focus is a quantitative deployment decision that can be adapted to the device’s resource budget.
Citation
Wen-Tzu Chang, Rui Fang, and Ming-Syan Chen. (2025). "SEAL: Secure and Efficient Adaptive Layering for On-Device Language Models with TEE." 2025 IEEE 11th International Conference on Edge Computing and Scalable Cloud, 138-143.