Weighted Semantic Models for Inference-Aware Deletion

Develop weighted dependency models for the semantics lens used in inference-aware deletion and design mechanisms that enforce deletion guarantees under such weighted semantics rather than binary dependency models.

Background

The paper introduces a semantics lens M to model how an adversary reasons about what remains inferable after deletion. Existing mechanisms largely rely on binary dependency models (e.g., functional or denial constraints) that either fully enable or disable an inference path.

The authors argue that realistic settings require graded or weighted dependencies to capture varying inference strengths across attributes and artifacts, and they explicitly call out the absence of such weighted semantics as an open area in the design space for deletion with bounded leakage.