TY - JOUR AU - Dulhanty, Chris AU - Wong, Alexander PY - 2020/01/02 Y2 - 2024/03/28 TI - Investigating the Impact of Inclusion in Face Recognition Training Data JF - Journal of Computational Vision and Imaging Systems JA - J. Comp. Vis. Imag. Sys. VL - 5 IS - 1 SE - Articles DO - UR - https://openjournals.uwaterloo.ca/index.php/vsl/article/view/1657 SP - 1 AB - <p>Modern face recognition systems leverage datasets containing im-<br>ages of hundreds of thousands of individuals’ faces. Recently, there<br>has been significant public scrutiny into the privacy implications of<br>large-scale training datasets such as MS-Celeb-1M, as many peo-<br>ple are uncomfortable with their face being used to train dual-use<br>technologies that can enable mass surveillance. However, the im-<br>pact of an individual’s inclusion in training data on a derived sys-<br>tem’s ability to recognize them has not previously been studied. In<br>this work, we audit ArcFace, a state-of-the-art, open-source face<br>recognition system, in a large-scale face identification experiment.<br>We find Rank-1 identification accuracy of 79.71% for individuals<br>present in training data and 75.73% for those not present. These re-<br>sults demonstrate that modern face recognition systems work bet-<br>ter for individuals they are trained on, which has serious privacy<br>implications as all large-scale, open-source training datasets do not<br>gather informed consent from individuals during their collection.</p> ER -