Source-Free Class-Balanced Pseudo-Label Adaptation for Large-Scale Cross-Day WiFi Transmitter Identification

Radio-frequency (RF) fingerprint distributions vary across collection dates, and changes in propagation conditions and hardware states can degrade transmitter-identification performance. To address this issue, we propose Source-Free Class-Balanced Pseudo-Label Adaptation (SF-CBPA) for cross-day WiFi transmitter identification. SF-CBPA requires only an unlabeled target-day dataset and a pretrained source model. It recalibrates batch-normalization statistics, freezes the source classifier, and uses Sinkhorn-balanced pseudo-labeling with within-class confidence selection; an exponential-moving-average teacher and information maximization provide auxiliary stabilization. Experiments were carried out using the public WiSig-ManyTx subset with 140 transmitters and one fixed receiver. The first three collection days trained the source model, while Day 4 provided 25 unlabeled adaptation samples and 25 independent test samples per class. Across three random seeds, SF-CBPA achieved 63.80 ± 0.90% accuracy and 0.6330 ± 0.0121 Macro-F1. A protocol-aligned SHOT implementation achieved 61.72 ± 1.95% accuracy and 0.5963 ± 0.0285 Macro-F1, while source-only and TENT-style achieved 57.93 ± 0.58% and 58.58 ± 0.43% accuracy, respectively. Direct analysis shows that Sinkhorn reduced the hard pseudo-label count range from 1-54 to 13-35 and the class-count standard deviation from 11.75 to 3.42. Component-isolated ablation identifies Sinkhorn balancing as the principal contributor under seed 123; the independent benefits of EMA and the auxiliary information-maximization term are small or inconsistent. These conclusions are restricted to the balanced, closed-set, fixed-receiver protocol evaluated here.