Adaptive federated few-shot feature learning with prototype rectification

Targeting to produce new features from limited data, few-shot feature generation approaches have attracted extensive attention and successfully mitigated the high cost of acquiring sufficient data. However, two main challenges remain underexplored among existing few-shot feature generation methods, namely the distribution gaps between base and novel classes, and the gradual tightening of data privacy. In order to ameliorate the performance drop induced by the distribution gap and alleviate the laborious cost of collecting massive data, in this paper, we propose a novel few-shot feature generation model that integrates domain alignment, prototype rectification, and federated learning into a unified framework. Concretely, the distance between across different classes is explicitly shrunk via domain alignment, facilitating more precise and reliable feature generation. Additionally, we develop prototype correction to reduce the intra-class discrepancy and make samples from the same class more clustered. Such that, the negative effects of the boundary samples are eliminated and thus boost the model performance. Finally, we combine our few-shot feature generation with the federated framework to protect data privacy and propose an adaptive federated scheme to provide customized services for individual clients. Extensive experiments are performed on three standard benchmark datasets to evaluate the effectiveness and superiority of our proposed method. The results consistently demonstrate that our proposed model gains substantial performance boosts and achieves state-of-the-art performance on the few-shot tasks.