Wi-Fi is no longer just for internet access. A new system named WhoFi uses how the human body alters radio waves to identify a person with up to 95.5 percent rank 1 accuracy, according to a new study.
It runs on consumer grade gear and still works in poor light. The approach relies on radio measurements, not cameras.
How the WhoFi system works
The idea is simple to state but technical under the hood. A WhoFi deployment captures Channel State Information (CSI), the fine-grained description of how a Wi-Fi signal changes as it moves through a room and around a person.
It then turns those changes into a compact biometric signature that is unique to that individual.
The work comes from Danilo Avola, of the Sapienza University of Rome, whose team built and tested the pipeline on public data.
The team reports that Wi-Fi is not just a stand-in for cameras, but offers different strengths that visual systems lack.
CSI is a matrix of amplitudes and phases across antennas and subcarriers. In plain terms, it captures tiny differences in how radio energy arrives at the receiver, which encode body shape and motion.
Those CSI sequences feed a deep network that learns a person-specific embedding. The best results came from a Transformer encoder that excels at long-range temporal patterns in the signal.
What the data say
The researchers evaluated WhoFi on the NTU-Fi Human ID benchmark, which contains recordings of 14 subjects performing short walks under different clothing conditions.
The NTU-Fi Human ID dataset includes measurements captured with higher resolution CSI tools that expose many subcarriers, allowing finer distinctions between people.
WhoFi’s top line result is a 95.5 percent rank 1 identification rate, with a mean average precision of 88.4 percent on the test split.
That score came from the Transformer configuration that was trained on amplitude sequences and validated against a held-out set of test data.
The hardware was not unusual. The paper documents tests using two TP-Link N750 routers, one transmitter and one receiver, recording 114 subcarriers per antenna pair and roughly 2,000 packets per sample.
The sample size is small, which matters because biometric systems can drift when scaled. The authors acknowledge that training stability and overfitting are real risks when models get deeper than necessary.
How WhoFi is different
Early attempts tied Wi-Fi to cameras to get the job done. In 2020, a camera Wi-Fi fusion system called EyeFi demonstrated about 75 percent accuracy in identifying people during live tests, when group sizes ranged from two to ten.
WhoFi goes a different route. It removes cameras from the loop and learns directly from CSI dynamics, which travel with a person across spaces.
The change is not just academic. Removing cameras avoids face capture, reduces sensitivity to clothing variation, and uses infrastructure that already exists in homes and offices.
Why it matters
The promise of radio based sensing is not new. A decade ago, researchers at MIT showed that low bandwidth Wi-Fi could track people through walls, establishing that consumer frequencies propagate through common barriers.
WhoFi rides that physics. It is insensitive to lighting, it can operate when a person is not in direct line of sight, and it can keep working when cameras would be occluded.
The system is also quiet in operation. There is no need to ask users to wear anything or carry a device, which changes the conversation about consent.
Wi-Fi routers send data across many narrow frequency slices called subcarriers.
When a person stands or walks between a transmitter and receiver, the amplitudes and phases on those slices shift in ways that correlate with their body and gait.
WhoFi ingests those time series and uses an encoder to create a fixed length vector that represents that person.
The encoder’s output is normalized, then matched against a gallery to see if there is a close neighbor from the same individual.
Training leans on in batch negatives. Each training batch pairs queries and galleries so the model learns to push mismatched people apart while pulling the same person’s signatures together.
Wi-Fi and WhoFi
The NTU-Fi benchmark is controlled, with short walks inside a defined area and set clothing conditions. That reduces real world noise from crowds, variable furniture, and reflections that can interfere in busy environments.
Sample diversity is another constraint. Fourteen subjects do not capture the range of body types, mobility aids, and cultural garments found in daily life.
“Wi-Fi signals offer several advantages over camera-based approaches: they are not affected by illumination, they can penetrate walls and occlusions, and most importantly, they offer a privacy preserving mechanism for sensing,” wrote Avola.
The paper’s conclusion stresses that the Transformer encoder was both accurate and efficient in this setting. The authors also note that common preprocessing steps, like amplitude filtering, did not always help.
Security and privacy questions
Accuracy numbers excite engineers, but deployment raises policy issues. A store could in principle use this technique to ping returning customers without asking for permission.
Law enforcement and regulators have a stake as well. Radio based identifiers might bypass laws written with cameras and faces in mind, which means legal frameworks will need review.
Wi-Fi operates on a shared spectrum, which makes sensing cheap to scale. That lowers the barrier for third parties who want to track presence and movement across access points.
The flip side is that any re-identification tool needs a gallery. Without a reference signature, a system can say the same person appeared twice, but it cannot assign a civil identity.
What happens with WhoFi next
There are benign uses. Hospitals might want fall detection in dark rooms, and home routers already ship with motion sensing features that rely on CSI.
Industrial safety is another candidate. In zones where cameras are banned, radio could watch for entry violations and trigger alarms.
The WhoFi team kept the system purely academic so far. They trained with public data, documented every step, and compared encoders in a reproducible way.
Future validation will need larger cohorts, varied buildings, and longer time gaps to test stability. It will also need clearer rules on consent and retention.
The study is published in arXiv.
—–
Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates.
Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.
—–
