Investigating the Impact of Inclusion in Face Recognition Training Data on Individual Face Identification

Training data inclusion provides a modest but meaningful performance advantage: Individuals whose faces were included in the training data achieved 79.71% face identification accuracy compared to 75.73% for those not included—a difference of approximately 4 percentage points in a large-scale test with over one million distractor images.

Better recognition of included individuals raises privacy concerns: The finding that face recognition systems work better for people in their training datasets creates a troubling asymmetry: people whose faces were used to train these systems without consent are more easily identified by them, potentially enabling more effective surveillance of those individuals.

The performance gap is "modest" but still significant: While the researchers describe the accuracy difference as modest, a 4-percentage-point improvement in a system already performing at ~80% accuracy represents a meaningful enhancement in real-world identification capabilities, particularly for mass surveillance applications.

This is the first empirical study of its kind: The paper addresses a previously unstudied question about how training data inclusion affects individual recognition performance, filling an important gap in understanding the practical implications of using people's faces to train face recognition systems.

Dual-use technology concerns are amplified: The findings underscore why public concern about face datasets like MS-Celeb-1M and MegaFace is justified—people included in training data face both privacy violations (unauthorized use of their image) and practical disadvantages (easier identification by surveillance systems).