Towards robust adaptive federated learning: unsupervised multidomain face recognition with Greater Cane Rat Algorithm and weighted fused extreme learning machine

Towards robust adaptive federated learning: unsupervised multidomain face recognition with Greater Cane Rat Algorithm and weighted fused extreme learning machine

연구 분야: Artificial Intelligence

학회: Iran Journal of Computer Science

Despite recent advances in computer vision in posture, orientation, and viewing mode of photos or videos that affect the device performance, facial expression cognition knowledge still faces challenges from some perspectives. Many state-of-the-art domain generalization (DG) techniques become unsuccessful in the Federated Learning (FL) situation because they need data centralization from multiple domains during training. In light of this, our work suggests a novel federated learning paradigm for post-exercise unsupervised multidomain face recognition. This research proposes a predictive model based on a personalized federated learning approach that integrates deep learning with an adaptive federated learning framework. The Greater Cane Rat Algorithm (GCRA) is employed to determine the optimal weight of the adaptive federated learning model. The flexibility of GCRA is controlled by a single adjustable parameter, which reflects whether it is mating season. During exploration, rats leave their shelters to forage and create trails, with the dominant male following these routes. Other rats adjust their positions based on this learned information, while males distinguish themselves during the breeding season. To effectively design the GCRA and optimize its tasks, rats’ intelligent foraging behaviors and mating season strategies are precisely modeled. Furthermore, a fused weighted deep extreme machine learning technique is incorporated to enhance multifacial expression recognition, further improving the accuracy and robustness of the proposed system. The experimental results in multidomain face recognition tasks demonstrate that the proposed system outperforms the benchmark approach, achieving an improvement of 12% in the F1 score. This enhancement is validated on multiple multidomain facial expression datasets, highlighting the effectiveness of the proposed method in handling diverse facial variations.