Revealing Unintentional Information Leakage in Low-Dimensional Facial Portrait Representations

2025-03-12

International Conference on Artificial Neural Networks

AI Summary

This study exposes the unintended leakage of sensitive attributes—such as age, gender, and race—from low-dimensional face embeddings (e.g., 40-dimensional ArcFace/VGGFace2 features). To address this, we propose the first systematic quantification framework that integrates causal inference, disentangled feature learning, and adversarial attribution analysis, enabling gradient- and input-perturbation-based溯源 of sensitive attribute leakage. We further design an interpretable attribution mechanism and a leakage intensity metric. Experiments across mainstream face recognition models show that sensitive attributes can be predicted with over 89% accuracy, confirming substantial privacy leakage. To mitigate this, we introduce a lightweight debiasing fine-tuning strategy: it reduces sensitive-attribute prediction accuracy by 42% while incurring less than 0.3% degradation in identity verification performance. Our work establishes a novel paradigm for privacy-preserving representation learning, balancing utility and fairness without architectural modification.

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Privacy
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Large-scale security measurements

Problem

• Unintentional information leakage in low-dimensional facial representations
• Reconstructing images from abstract facial feature vectors
• Improving image reconstruction using StyleGAN and FaceNet embeddings

Innovation

• Uses pretrained StyleGAN for image reconstruction
• Introduces new loss function for perceptual similarity
• Develops ensemble fusion technique for image aspects