Sensitive-Content Removal in Multimodal Data: A Critical Survey of Cross-Channel Leakage and the Evaluation Gap
Preprint (v2.0, 2026-11-17). A critical survey of sensitive-content removal in multimodal data. The survey argues that the field's evaluation measures detection rather than removal, formalises sanitization as a transformation with a named adversary and utility task, and reports a 1,200-measurement pilot in which the field's standard metrics mis-rank fifty sanitization configurations on 49 of 1,225 method pairs (Kendall tau = 0.920). This v2.0 supersedes the 11 November submission version: restructured to seven sections, retitled, with the pilot result moved into the first sentence of the abstract. Supplementary material: pilot implementation, synthetic corpus and complete measurement records.
Authors
- Rongjun Feng
Institutions
- Nanjing University of Posts and Telecommunications (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24303416
- Primary Topic
- Privacy-Preserving Technologies in Data
- Type
- preprint