Membership Inference in a Handwriting Recogniser Trained on a Community Archive
A research poster from the School of Underlying Conditions trains a handwriting recogniser on seventeen years of a queer support collective's letters and newsletters and tests which of its 1,388 name strings a membership-inference attack can recover. Leakage concentrates in 23 names written in full on newsletter sign-offs; names written as initials stay at chance, while differentially private training protects the organisers only by degrading recognition of the archive's rarest hands.
| Authors | Brynjar Obiora , Mirela Hanke , Iben Chikere |
|---|---|
| School | School of Underlying Conditions |
| Output type | Research poster |
| Published | |
| DOI | 10.5555/slop.juvgwh |
| Pages | 1 |
| Cited by | 1 output |
| Version | 1.0 |
| Licence | CC BY 4.0 |
Cites
Cited by
Cite as
@misc{slop_juvgwh,
author = {Brynjar Obiora and Mirela Hanke and Iben Chikere},
title = {The Names That Were Signed: Membership Inference in a Handwriting Recogniser Trained on a Community Archive},
year = {2026},
publisher = {Slop University},
doi = {10.5555/slop.juvgwh},
url = {https://slop.university/outputs/slop-poster-membership-inference-leakage-juvgwh/},
version = {1.0},
note = {Research poster},
}