Adaptive Confidence-Driven Sample Selection with Alignment Loss for Efficient Federated Active Learning

Abstract

Federated Active Learning (FAL) aims to reduce labeling costs by selecting the most informative samples for annotation while preserving strong model performance. However, current FAL methods struggle with the challenge of client heterogeneity in non-IID environments, often leading to suboptimal sample selection and increased computational burden due to repeated inference on unlabeled data. To address these issues, we propose a novel approach called Adaptive Confidence-Driven Sample Selection (ACdSS), which combines the CHASe strategy with Alignment Loss to improve sample selection efficiency. ACdSS leverages dynamic confidence scores to assess uncertainty in unlabeled data and identifies samples with fluctuating predictions between training rounds, which are more likely to be near the decision boundary. This approach balances exploration and exploitation by adapting the sample selection strategy throughout the training process, minimizing inference costs while maintaining high selection effectiveness. Experimental results on benchmark datasets demonstrate that ACdSS outperforms existing FAL techniques, delivering superior accuracy with reduced annotation costs. This framework offers a practical solution for managing client heterogeneity in FAL and is well-suited for real-world federated learning systems, offering a significant reduction in computational overhead while improving model performance.