Original title: Source-Free Class-Balanced Pseudo-Label Adaptation for Large-Scale Cross-Day WiFi Transmitter Identification
Yuheng Yang, H ZHANG, Xiaoya Wang, Mingdi Li +1 · September 2026
Every WiFi transmitter emits a unique radio-frequency fingerprint. Over time, shifting weather, room conditions, and warming hardware alter these signals. Because of these daily changes, models trained to identify transmitters on one day often struggle on a new day. Researchers built a technique called Source-Free Class-Balanced Pseudo-Label Adaptation, or SF-CBPA. It updates the system using only the old model and new, unlabeled signals. It freezes parts of the model and recalibrates its tracking statistics. To prevent the model from guessing the same common transmitters repeatedly, it uses an algorithm called Sinkhorn balancing. This mechanism spreads guessed labels evenly across the devices and keeps only the most confident predictions. In tests, SF-CBPA achieved 63.80 ± 0.90% accuracy, outperforming older adaptation methods. Sinkhorn balancing provided the main performance boost by keeping label counts balanced across classes.
You are a teacher who knows each student's handwriting from September. In October, some students change their writing style, so you can no longer tell who wrote what. You have a pile of new homework with no names, but you know the students are the same. Here, the teacher is the identification system, the handwriting is the radio fingerprint, and the new homework is the unlabeled data from a different day. The method assigns each homework to a student while making sure no student gets too many or too few assignments.
WiFi networks use the unique radio fingerprint of each device to identify it. But those fingerprints change from day to day due to weather and hardware, causing identification to fail. This method lets the system adapt to a new day using only unlabeled data, keeping security and monitoring working without costly relabeling.
The paper only evaluates on a balanced, closed-set, fixed-receiver protocol. It does not test on imbalanced data, unknown devices, or multiple receivers. The improvement over baselines is modest, and only the Sinkhorn balancing component consistently helps; other components show small or inconsistent effects.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Source-Free Class-Balanced Pseudo-Label Adaptation for Large-Scale Cross-Day WiFi Transmitter Identification — Yuheng Yang, H ZHANG, Xiaoya Wang et al. — Sensors — doi:10.3390/s26185975
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.