Adaptive Identity Leakage in Motion Anonymization: A CMU Benchmark and Evaluation Framework
Abstract
Motion anonymization must suppress identity information that remains usable after an attacker adapts to the released signal. This paper examines that requirement through a retrospective analysis of a reported CMU motion-capture benchmark and an explicit framework for interpreting its evidence. The benchmark contains 2,657 six-second head-and-hand trajectory segments from 27 subjects and compares a bounded adversarial transformation, AURA, with temporal downsampling, head-only disclosure, Gaussian perturbation, and raw motion. Across the reported seed–budget aggregates, random-forest top-1 identification increases from 0.111 to 0.532 for downsampling and from 0.148 to 0.571 for head-only disclosure after attacker retraining. AURA yields 0.593 against the adaptive random forest, compared with 0.598 for raw motion. These are descriptive results: exact enrollment–test overlap affects 56/261 and 178/240 pooled test segments at the two larger budgets, and the execution package is unavailable for independent replication. We derive sharp bounds for clean-subset accuracy under a specified contamination mixture, distinguish nominal window duration from distinct enrollment exposure, and establish why small or invertible transformations do not alone imply identity protection. We also define a trial-disjoint replication protocol, a common motion-quality measurement contract, and the requirements for a differential-privacy claim. The resulting framework separates evidence of attacker distribution shift from evidence of reduced identity information while preserving the benchmark’s unresolved empirical limitations.