Generating adversarial samples for emotion recognition based on diffusion model

• 1、School of Cyberspace Security, Beijing University of Posts and Telecommunications, Beijing, 100876

Abstract：Existing adversarial example generation methods for Facial Expression Recognition (FER) systems generally suffer from severe visual distortion and unnatural expressions, making it difficult to achieve both high attackability and high concealment in real-world scenarios. To address the issue of low image quality caused by traditional pixel-level perturbations, this paper proposes an adversarial example generation method based on a diffusion model. This method transfers perturbation optimization to the high-level semantic space, applies controlled perturbations on the semantic encoding extracted by a diffusion autoencoder, and combines multi-objective loss functions such as emotional attack and action unit (AU) rationality to guide the evolution of emotional representations towards the target category, generating high-fidelity and physiologically plausible adversarial examples. Finally, simulation experiments demonstrate that the proposed method significantly outperforms existing advanced baseline methods in multiple key metrics, including attack success rate and visual quality. The generated adversarial examples can effectively deceive mainstream FER models while maintaining high visual naturalness, providing strong technical support for the active protection of facial expression privacy.

Keywords： Emotion recognition adversarial example generation diffusion model