Cross-domain facial expression recognition based on adversarial attack fine-tuning learning

Expression recognition is important in artificial intelligence research and has broad application prospects in medical care and transportation. However, owing to differences in lighting, culture, ethnicity, etc., cross-domain expression recognition is difficult. Analysis of the mechanism related to graph error discrimination is critical for improving cross-domain expression recognition performance. However, effective methods for expression error prediction and model performance analysis are lacking. In this study, an adversarial attack method is proposed to realise the analysis and adversarial attack fine-tuning learning is used to improve cross-domain expression recognition. To implement this method, the adversarial attack dataset must include domain differences in brightness, contrast, grayscale, Gaussian perturbation, and geometric perturbation. The critical graph feature fusion network, constructed using a residual network, local features, and prior graph-learning techniques, guarantees the realisation of this method. An adversarial attack expression recognition experiment revealed the influence of the attribute parameters of the database on graph error discrimination and illustrated that finetuning learning can prevent graph error discrimination. In the adversarial attack fine-tuning cross-domain expression recognition experiment, an average improvement of 7.09% in the recognition accuracy was achieved on the four datasets, even without introducing any information about the target dataset. It can be concluded that the model achieved significant improvement in feature extraction performance and expression recognition range. Adversarial attack fine-tuning learning was integrated into the two best-performing cross-domain recognition methods, achieving an average accuracy improvement of no less than 6.93% on four datasets, demonstrating the universality of the proposed method.