Retaining Emotions, Removing Identity: Valence–Arousal GuidedDe-Identification for Educational Applications

Abstract
The integration of artificial intelligence (AI) for multimodal data analysis has enabled deep learning models to analyze students’ facial expressions for emotion recognition in classroom settings. However, collecting facial data from minors raises substantial privacy and ethical concerns, necessitating privacy-preserving deidentification techniques that retain affective information. Existing methods primarily target identity suppression but often fail to preserve fine-grained emotional expressions, limiting applicability in affective computing. We present a privacy-preserving face deidentification method that replaces real identities with synthetic faces selected via valence–arousal guided identity replacement in the StyleGAN latent space. This approach ensures de-identified faces retain the original emotional expressions while suppressing biometric identity. We evaluate our algorithm on a real-world classroom dataset of 40 middle school students and the publicly available DAiSEE dataset, achieving privacy preservation accuracy of 89.97% (Classroom) and 89.94% (DAiSEE), significantly outperforming conventional anonymization techniques. Additionally, the method maintains high expression fidelity, with 92.4% manual verification agreement and low Facial Action Unit (FAU) loss. These findings establish an efficient, privacy-compliant de-identification framework for affect-sensitive AI supporting classroom analytics and educational research.