Physical-level privacy-preserving face recognition via optically encoded single-pixel measurements

Effective data privacy protection in face recognition has become increasingly critical, particularly in the era of generative models and deepfake technologies. To achieve privacy protection at the physical level, we propose an image-free face recognition method based on single-pixel measurements. It integrates physical-domain optical encoding with a neural network-based back-end, which operates directly on one-dimensional light intensity signals. This enables end-to-end recognition without reconstructing any visual images throughout the entire pipeline. A learnable convolutional encoder based on the straight-through estimator models the optical modulation process, while a dual-branch inverse feature expansion network improves recognition performance on compressed measurements. Experiments on both synthetic and real-world data demonstrate that our method achieves 95.56 % recognition accuracy at a 1 % sampling ratio, outperforming existing approaches in terms of sampling ratio, recognition accuracy, and privacy-preserving capability. This work facilitates the application of image-free sensing techniques in privacy-sensitive scenarios.