Abstract
Modern face recognition systems leverage datasets containing images of hundreds of thousands of individuals’ faces. Recently, there has been significant public scrutiny into the privacy implications of large-scale training datasets such as MS-Celeb-1M, as many people are uncomfortable with their face being used to train dual-use technologies that can enable mass surveillance. However, the impact of an individual’s inclusion in training data on a derived system’s ability to recognize them has not previously been studied. In this work, we audit ArcFace, a state-of-the-art, open-source face recognition system, in a large-scale face identification experiment. We find Rank-1 identification accuracy of 79.71% for individuals present in training data and 75.73% for those not present. These results demonstrate that modern face recognition systems work better for individuals they are trained on, which has serious privacy implications as all large-scale, open-source training datasets do not gather informed consent from individuals during their collection.
Introduction
Modern Face Recognition (FR) systems rely on Deep Convolutional Neural Networks (DCNNs) trained on massive datasets containing millions of images. While these systems have reached incredible levels of accuracy, they have also sparked intense ethical debates regarding mass surveillance and privacy. A major point of contention is the use of datasets like MS-Celeb-1M, which were scraped from the internet without the explicit consent of the individuals pictured. While we know these datasets improve FR systems generally, this research investigates a more specific, personal question: Does including your face in the training data make the system specifically better at identifying you? If so, being "included" in a dataset represents a specific privacy liability rather than just a general contribution to model performance.
Methodology
The authors performed a large-scale audit of ArcFace, a leading open-source FR model, framing the experiment as a closed-set face identification task. The model was trained on a cleaned version of the MS-Celeb-1M (MS1M) dataset. For a given "probe" image, the system must find the correct identity within a "gallery" of 1 million "distractor" images sourced from the MegaFace Challenge 1.
To measure the impact of inclusion, two distinct probe sets were created:
- Training Probe Set: 1,000 identities present in the model's original training data (MS1M).
- Novel Probe Set: 1,000 identities the model had never seen during training.
The pipeline utilized MTCNN for face detection and alignment, followed by the generation of 512-dimensional feature representations using ArcFace. Identification was determined by calculating the L2 distance between the probe embedding and the gallery embeddings.
Datasets
The study utilized several cornerstone datasets in the FR field:
- MS-Celeb-1M (MS1M): Used for initial training, containing ~5.2 million images of ~93k identities.
- VGGFace2: Used to source probe images. Matching names between VGGFace2 and MS1M allowed the authors to distinguish between "known" and "novel" identities.
- MegaFace Challenge 1: Provided the "distractor" gallery of 1 million images (~690k identities), creating a challenging "needle-in-a-haystack" scenario.
Results and Key Insights
The results showed a consistent performance gap across all metrics, confirming that the model is significantly better at identifying individuals it encountered during training.
As shown in the table, the system achieved a Rank-1 accuracy of 79.71% for the training set versus 75.73% for the novel set—a 3.98% "inclusion bonus." This gap persisted at Rank-10 and Rank-100, suggesting the effect is robust. This finding suggests that FR models do not just learn general facial features; they appear to optimize specifically for (or "memorize") the identities they are trained on. This provides a technical, quantitative metric for privacy loss: the delta in identification accuracy between being in versus out of a training set.
Discussion and Conclusion
The performance boost for training identities highlights a "dual-use dilemma." If a population's data is scraped to train a system, that population becomes significantly more vulnerable to automated tracking by that specific system. This is particularly concerning given the lack of informed consent in large-scale dataset collection.
While this study focuses on ArcFace and MS1M, future work should investigate if different architectures or loss functions exhibit similar gaps. Additionally, exploring whether certain demographic groups experience a larger "inclusion bonus" could provide further insights into algorithmic fairness and privacy risks. These findings underscore the need for more rigorous consent frameworks in dataset collection.
Summary updated