Vector Order Shuffling
A privacy defense technique that keeps vector contents unchanged and only randomizes their order in the index, attempting to hide the spatial layout of patches. It can be reversed by models specifically trained to recover the order, and is therefore regarded as an illusion of security.
Timeline (last 90 days)
Both defenses, token pooling and vector order shuffling, can suppress text recall to about 8%.
Against order shuffling, a model specifically trained to recover order can raise the proportion of "source page ranked first" from 3.8% to 93.5%, almost returning to the level without defense.
Although the pooled version can still be inverted, text recall remains at a low level, and the actual leakage risk is clearly lower than with order shuffling.
The paper explicitly points out that a defense relying solely on obfuscating vector order can be reversed by a learning model of moderate scale.
Defense design should prioritize disrupting patch-level semantic granularity, rather than merely hiding the arrangement; combining the two still requires careful evaluation.
All Facts (5)
Both defenses, token pooling and vector order shuffling, can suppress text recall to about 8%.
50%UnverifiedAgainst order shuffling, a model specifically trained to recover order can raise the proportion of "source page ranked first" from 3.8% to 93.5%, almost returning to the level without defense.
50%UnverifiedAlthough the pooled version can still be inverted, text recall remains at a low level, and the actual leakage risk is clearly lower than with order shuffling.
50%UnverifiedThe paper explicitly points out that a defense relying solely on obfuscating vector order can be reversed by a learning model of moderate scale.
50%UnverifiedDefense design should prioritize disrupting patch-level semantic granularity, rather than merely hiding the arrangement; combining the two still requires careful evaluation.
50%