LipCache: Certified Semantic Caching for Edge Image Classification
A new system called LipCache has been created to reduce the costs of using vision models in edge computing while maintaining dependability. Detailed in a paper on arXiv (2608.13144), it addresses issues with current semantic caching methods that rely on similarity thresholds, which can cause missed misclassifications at decision boundaries. LipCache includes GuardNet, a lightweight network that converts inputs into a lower-dimensional feature space under a Lipschitz constraint. It calculates a certified reuse radius for each sample based on the local classification margin and the spectral norm. Cached results are reused only if the query feature falls within this radius, ensuring accuracy. The original model, MainNet, remains unchanged, aiming for efficient edge vision services.
Key facts
- LipCache is a certified semantic caching framework for image classification.
- It addresses silent misclassifications near decision boundaries in existing semantic caching methods.
- The framework introduces GuardNet, a lightweight network mapping inputs to a low-dimensional feature space under a Lipschitz constraint.
- It computes a per-sample certified reuse radius from the local classification margin and spectral norm of the classification head.
- Cached results are reused only when the query feature falls within the certified radius.
- The existing deployed main model (MainNet) is not modified.
- The paper is available on arXiv with identifier 2608.13144.
- The announcement type is cross, suggesting prior or parallel publication.
Entities
Institutions
- arXiv