Abstract
Standard federated averaging gives clients with larger training datasets greater weight. How reducing these weight differences affects overall and client-level short message service (SMS) spam detection remains unclear. We compare nine federated methods using logistic regression on 5023 ExAIS messages from 20 clients, defined by source files. The methods include equal weighting, weighting proportional to training size, and square-root weighting to reduce differences in aggregation weights. Evaluation repeats two-fold cross-validation five times, assigning identical normalized messages entirely to the training or testing set in each split. With class-weighted loss, three local epochs per round, and a threshold of 0.5, square-root weighting has the highest mean pooled F1 (0.8150), calculated from all test messages together within each fold. However, none of the four primary comparisons is significant after multiple-testing correction. Equal-client weighting has the highest observed average client F1 and mean F1 for the lowest-performing quarter of clients. Thresholds selected on inner validation data reduce the square-root weighting’s advantage over class-weighted proportional aggregation from 0.0112 to 0.0010. With equal local update counts per client and round, proportional weighting ranks first on pooled F1, followed by square-root and equal-client weighting. In this equal-step analysis, only the square-root versus equal-client contrast is significant after correction for the two post hoc comparisons. On an external SMS corpus, equal-client weighting ranks first among the three rules, with a mean pooled F1 of 0.3960. No aggregation rule performs best across all evaluated settings, and rankings depend on local training, classification thresholds, and whether performance is assessed across messages or clients.