Source-linked AI summary
WiSig: A Large-Scale WiFi Signal Dataset for Receiver and Channel Agnostic RF Fingerprinting
Samer Hanna, Samurdhi Karunaratne, Danijela Cabric
TL;DR
RF fingerprinting must distinguish transmitter-specific variability from impairments introduced by channels and receivers, but suitable broad data are limited. WiSig provides a large, publicly accessible WiFi dataset with processing resources, and its preliminary evaluation shows that changing days or receivers can degrade classifier performance. The dataset supports development and evaluation of more channel- and receiver-agnostic approaches.
Problem
Channel and receiver fingerprints confound data-driven transmitter identification, while practical systems need to operate across changing channels, prolonged periods, and multiple receivers.
Method
WiSig constructs and publicly distributes a large WiFi dataset with raw and processed signals, pre-packaged subsets, scripts, and examples.
Results
WiSig contains 10 million WiFi packets from 174 off-the-shelf transmitters and 41 USRP receivers across four captures over a month, with 9.97 million identification signals reported for the full configuration.
Takeaways & Limitations
WiSig supports studying how transmitter-identification performance changes with transmitters, receivers, days, signals, and equalization, while exposing open problems for more robust systems.
Abstract
from arXiv · showhide
RF fingerprinting leverages circuit-level variability of transmitters to identify them using signals they send. Signals used for identification are impacted by a wireless channel and receiver circuitry, creating additional impairments that can confuse transmitter identification. Eliminating these impairments or just evaluating them, requires data captured over a prolonged period of time, using many spatially separated transmitters and receivers. In this paper, we present WiSig; a large scale WiFi dataset containing 10 million packets captured from 174 off-the-shelf WiFi transmitters and 41 USRP receivers over 4 captures spanning a month. WiSig is publicly available, not just as raw captures, but as conveniently pre-processed subsets of limited size, along with the scripts and examples. A preliminary evaluation performed using WiSig shows that changing receivers, or using signals captured on a different day can significantly degrade a trained classifier's performance. While capturing data over more days or more receivers limits the degradation, it is not always feasible and novel data-driven approaches are needed. WiSig provides the data to develop and evaluate these approaches towards channel and receiver agnostic transmitter fingerprinting.
I. INTRODUCTION
RF fingerprinting uses transmitter-specific hardware variability, but wireless channels and receiver circuitry introduce confounding fingerprints. WiSig addresses the resulting data gap with a large, public WiFi dataset spanning many transmitters, receivers, and capture days.
- Motivation: Manufacturing variability gives transmitters distinct RF fingerprints, even when same-model devices send identical waveforms.These differences can support transmitter identification without transmitter-side overhead.
- Motivation: Channel and receiver fingerprints are embedded in captured signals, confounding data-driven transmitter identification.A reported channel change reduced classifier accuracy from 85% to 9%.
- Motivation: Practical systems may involve hundreds of transmitters, changing channels over prolonged periods, and many receivers, motivating channel- and receiver-agnostic fingerprinting.Training a separate classifier for each receiver may not be practical.
- Related Work: Prior datasets generally used custom testbeds, were often inaccessible, considered at most 100 transmitters, or included only one receiver.The cited literature includes datasets for LoRa, ZigBee, and WiFi, with varying public availability and hardware choices.
- Contribution: WiSig contains 10 million WiFi packets from 174 off-the-shelf transmitters and 41 USRP receivers across four captures performed within a month.The dataset is presented as the largest publicly available dataset and the first described here to include such a large number of both transmitters and receivers.
- Contribution: WiSig includes raw and processed data, pre-packaged subsets, processing and replication scripts, and examples for studying transmitter, receiver, day, and signal-count effects.The paper also uses the dataset to identify open RF-fingerprinting problems.
II. DATASET OVERVIEW
WiSig captures WiFi transmissions in the Orbit testbed using many transmitters, USRP receivers, and capture days, then processes and packages the data for classifier use.
- Capture setup: WiSig records WiFi transmissions from nodes to an access point while USRP receivers continuously capture raw IQ samples.The WiFi access point establishes the link; the WiFi nodes are transmitters and the USRPs are receivers.
- Data preparation: The processing creates equalized and non-equalized identification signals from detected packets, and users can select packaged subsets by transmitter, receiver, day, and signal count.The dataset is supplied with scripts for moving between processing stages and with compact prepackaged subsets.
- Capture setup: The Orbit testbed uses a 20 × 20 grid of roof-mounted computers, with some nodes equipped with USRPs and additional USRPs in massive MIMO racks.The grid nodes are separated by about 1 m, and the dataset uses several WiFi module and USRP models.
- Capture schedule: Captures were performed on four days in March 2021, while transmitter and receiver availability varied across days because of access and hardware or software problems.Consequently, some transmitter-receiver pairs have no raw data on some days.
- Capture configuration: All transmitters use the same spoofed MAC and IP addresses to avoid exposing identity clues through network identifiers.The WiFi configuration used channel 13 at 2462 MHz with 20 MHz bandwidth.
IV. EXTRACTING IDENTIFICATION SIGNALS
WiSig extracts classifier-ready identification signals by detecting and isolating WiFi packets, then generating both unprocessed and equalized sample representations.
- Packet isolation: Packet processing detects relevant transmissions, removes irrelevant signals and access-point acknowledgments, and isolates WiFi packets.Detection compares signal magnitude over a window against a fixed threshold, while packet duration distinguishes WiFi packets from ACKs.
- Signal representations: The non-equalized identification signal uses the first 256 IQ samples from each isolated packet.These samples are retained with transmitter, receiver, and day labels.
- Signal representations: The equalized representation resamples packets to 20 Msps, detects the preamble start, estimates and corrects frequency offset, and applies MMSE channel equalization.The WiFi L-STF and L-LTF preambles support synchronization and channel estimation; packets with failed L-STF detection are discarded.
- Signal representations: The resulting datasets contain unprocessed preamble samples and equalized signals for use as identification-signal inputs.Both representations are built from the isolated packets.
V. DATASET ASSEMBLY
WiSig assembly analyzes uneven signal availability and provides selection methods for constructing subsets with specified transmitter, receiver, day, and signal-count requirements.
- Subset selection: WiSig users choose subsets by specifying transmitters, receivers, days, and the maximum number of signals to retain.The assembly flow begins from all identification signals and supports application-specific dataset sizes.
- Dataset analysis: 9.97 million identification signals were collected from 174 transmitters and 41 receivers over four days.Signal counts vary because of failures, WiFi MAC rate control, receiver nonsynchronization, and detection errors.
- Dataset analysis: Most transmitter-receiver-day counts are below 400, while some exceed 1000 and a few exceed 2000.The corresponding count visualization uses a logarithmic colormap for the first day and a histogram across all days.
- Dataset analysis: Large count imbalance can confuse training and produce misleading results, motivating subset selection to reduce imbalance.The required number of transmitters, receivers, or signals may also be smaller than the full dataset.
- Selection optimization: Subset selection chooses transmitter and receiver sets whose size and minimum per-pair signal count satisfy a specified coverage fraction.The variables N, M, and K are dependent, so only two can be specified while the third is bounded by the dataset.
- Selection optimization: A greedy heuristic selects transmitters and receivers for p=1, while a MILP formulation is used when p=0.9 because the heuristic is not acceptable.The p=0.9 MILP result can utilize a larger portion of the dataset.
C. Prepackaged Compact Subsets
WiSig provides four compact, prepackaged subsets that emphasize different combinations of transmitters, receivers, signals, or capture days.
- Subset designs: ManySig is designed to be balanced, providing 1000 signals for every transmitter-receiver pair across all four days.It emphasizes consistent signal availability across pairs and days.
- Subset designs: ManyTx emphasizes transmitter scale with 150 transmitters while allowing slight imbalance at p=0.9.ManyRx instead provides signals from 32 receivers with slight imbalance.
- Subset designs: SingleDay provides relatively many signals and transmitters for one day, supporting users who do not need to study day-related channel variation.ManyTx, ManyRx, and ManySig cover all four days.
- Download options: The compact subsets require downloads of only a few gigabytes, unlike the larger Full WiSig dataset.Table III reports the download sizes for the available dataset versions.
VI. USE CASES & OPEN PROBLEMS
WiSig is used to examine how receiver and day variation affect transmitter classification, while also identifying open problems in robust fingerprinting. The evaluation is deliberately preliminary and focuses on closed-set classification.
- Evaluation setup: The evaluation uses a fixed closed set of transmitters and a simple neural-network setup to expose open problems rather than solve them.The network has five convolutional layers followed by three dense layers and is trained with early stopping.
- Generalization across receivers: Receiver variation is studied with the ManyRx dataset by training on subsets of 32 receivers using equalized data from one day.The experiment evaluates classifiers on the same training receivers and on five different receivers.
- Generalization across days: The ManySig experiment varies the number of training days and compares equalized with non-equalized data on the same training days and a different day.Each number of training days receives separate neural networks for the two data treatments.
- Open problems: Using only one receiver may be impractical for deployment across arbitrary receivers, motivating receiver-fingerprint augmentation or receiver-independent features.WiSig provides data for developing and evaluating both approaches.
B. Generalization across Days
WiSig evaluates how transmitter classifiers generalize across capture days and how training-signal quantity and equalization affect performance. The results motivate data-efficient and channel-robust approaches because prolonged data collection is undesirable.
- Generalization across days: A classifier trained on one, two, or three days is tested on a held-out fourth day using one receiver from ManySig.Both equalized and non-equalized datasets are evaluated.
- Generalization across days: Channel estimates for one transmitter–receiver pair differ across March 1st and 8th despite per-packet power normalization.The estimates cover all 52 occupied WiFi subcarriers, with means and standard deviations computed over packets.
- Impact of number of training signals: Equalization significantly improves performance with few signals, while increasing signal count improves both approaches and narrows their gap.The experiment uses all four days and twelve receivers in ManySig, with matching conditions for training and testing.
- Open problems: With enough signals from all days and receivers, classifiers can counteract channel randomness, but prolonged capture is not always desirable.The paper therefore points to channel augmentation and architectures robust to channel variation as open directions.
D. Impact of Number of Transmitters
WiSig supports large-scale transmitter experiments and shows that classification becomes harder as the number of transmitters increases. Its many-transmitter and grid-based data also support open-set, lookup-based, and localization studies.
- Impact of number of transmitters: ManyTx contains signals from 150 transmitters, enabling transmitter-authorization evaluation at a large user scale.This addresses applications that may need to authenticate hundreds of users.
- Impact of number of transmitters: Around 80% accuracy is obtained with 10 transmitters using four days, 19 receivers, and 50 signals per transmitter, receiver, and day.The evaluation uses non-equalized data from all days and receivers.
- Impact of number of transmitters: Accuracy drops to about 53% with 150 transmitters under the same large-scale evaluation, highlighting the need for architectures that handle many transmitters.The problem becomes more challenging as the transmitter count increases.
- Practical formulations: Many transmitters enable open-set recognition and fingerprint-vector database lookup, which require separating training transmitters from unseen test transmitters.Open-set recognition can reject unauthorized transmitters unseen during training.
- Localization: WiSig’s transmitter–receiver grid also supports localization experiments based on average received baseband power across receivers.Example transmitters are plotted against receiver heatmaps in dBm.
- Conclusion: The conclusion reports that changing days, receivers, or limiting training signals harms performance, while more data or equalization improves generalization but may be infeasible.Further analysis is needed to characterize combined effects and data requirements.
APPENDIX
The appendix formulates transmitter and receiver selection as a mixed-integer linear program that satisfies data-availability requirements across days and both non-equalized and equalized signals. Its hierarchical objective first maximizes selected receivers, then maximizes the minimum qualifying capture count.
- MILP formulation: The MILP selects transmitter and receiver subsets that satisfy specified requirements across all days and both non-equalized and equalized data.The formulation uses a generalized version of the earlier selection problem and is provided with accompanying helper code.
- Requirements: The user specifies the number of transmitters, the capture threshold K, and a lower bound K_low that tolerates shortfalls below K.K_low ensures that every selected transmitter-receiver-day combination below K still meets the specified lower bound; the original formulation uses K_low = 0.
- Objective: The hierarchical objective first maximizes the number of selected receivers M = |R|, then maximizes the minimum capture count k without reducing |R|.An extremely large weight w_∞ implements the priority given to maximizing receiver coverage.
- Variables: Binary variables T and R indicate selected transmitters and receivers, while Y encodes their pairwise selection.The dataset contains 174 transmitters and 41 receivers, and Y(t, r) = T(t) & R(r).
- Variables: Binary constants encode whether capture counts meet threshold K, and binary variables Q and Qeq identify qualifying selections for non-equalized and equalized data.The corresponding slices q(d, r) collect Q(t, d, r) values over the selected transmitter set; Z is an integer variable bounded using U.