This paper addresses the challenge of detecting subtle data manipulations in federated learning marketplaces, focusing on synthetically generated datasets introduced by data providers to inflate their contributions’ value. We propose a novel approach incorporating enhanced monitoring features and sophisticated analysis of local model weight movements. Our method measures the moving distance of local model weights, applies time series anomaly detection, and analyzes cross-client correlations to distinguish between genuine learning divergence and manipulation attempts. Experiments in realistic scenarios, in which 100% of recall of data modification events were detected, demonstrate the effectiveness of our approach in identifying subtle data modifications that previous methods are miss. This work contributes to the security and trustworthiness of federated learning marketplaces by providing a robust defense against an emerging threat in collaborative AI development.