The idea is simple. Wireless signals bounce, bend, and shift when passing through the human body. These changes aren’t random. They follow patterns shaped by body size, movement, and internal structure. Even small differences, like how someone walks, affect the signal.
Each time a person moves through a signal path, they leave behind a unique trail of distortions. These trails can be measured. And, with the right system, they can be matched.
How the System Reads the Signal
The system is called WhoFi. It doesn’t rely on facial features or sound. Instead, it records Channel State Information, or CSI. This is a detailed snapshot of how a Wi-Fi signal changes across time, space, and frequency.
First, the raw signal is cleaned. Outliers are removed. Phase shifts caused by hardware delays are corrected. Once the data is stable, it's pushed into a neural network.
That model has two main parts. One compresses the cleaned signal into a shorter format. The other turns it into a signature. This signature doesn’t describe what someone looks like. It reflects how they affect the air around them.
It’s a set of numbers. That’s all. But those numbers can tell one person apart from another, even across different environments.
Test Setup and Subjects
To test the system, researchers collected data from 14 people. Each person walked through a Wi-Fi zone multiple times. They changed clothes. Sometimes they wore coats. Other times, backpacks. Even with those changes, the system still picked them out.
The equipment used wasn’t special. It came from off-the-shelf TP-Link routers. One router sent the signal. Another, with three antennas, received it. For every recording, over two thousand packets of data were collected. That added up to a lot of signal snapshots, plenty for analysis.
Only the amplitude values were used in training. No original waveform or raw phase matrix was needed. The system worked with what was publicly available.
Comparing Models for Pattern Recognition
Three types of learning models were tested. One used LSTM. Another used Bi-LSTM. The third used a Transformer. These models were trained to recognize signal patterns and link them back to individuals.
The Transformer model outperformed the others. It handled longer sequences better. It picked up on subtle variations that the others missed. Its top-ranked accuracy reached 95.5 percent.
The LSTM models worked too, just not as well. They struggled with long input sequences and needed more careful tuning. They showed decent performance but couldn’t match the precision of the Transformer.
Sequence Length Matters
Longer sequences gave the model more to work with. For the Transformer, that meant better accuracy. It was able to model long-range dependencies in signal data. This helped it detect repeating structures and patterns tied to a person’s walk.
LSTM models didn’t benefit as much. In some cases, longer inputs even made them worse. That’s likely due to vanishing gradients and memory limits. Still, their performance stayed within usable range.
The lesson was clear. When it comes to signal-based identity tracking, more data helps, but only if the model can handle it.
Data Augmentation Improves Flexibility
To build flexibility into the system, researchers trained it with slightly altered input data. They added noise. They shifted time steps forward or backward. They adjusted signal strength slightly.
These changes made the model less sensitive to noise and better at generalizing. LSTM-based models saw the biggest improvement. The Transformer was already strong, but the added variation made it more resilient to real-world inconsistencies.
Interestingly, too much filtering hurt performance. Removing all irregularities sometimes erased useful information. Not all noise is bad. Some of it carries subtle identity cues.
No Names, Just Patterns
The system doesn’t store names or faces. It doesn’t label people. It only remembers signal patterns and compares them later. If two sessions match, it assumes it’s the same person.
This has clear technical benefits. It avoids collecting direct personal data. But it also introduces concerns. The system can track individuals across time and location, even if it doesn’t know who they are.
That tracking can happen without consent. And it can work in places where cameras fail, through walls, in darkness, or in crowded spaces. That makes it useful, but also sensitive.
Research Only, For Now
The WhoFi system is not being deployed. It’s not for sale. The project was developed for research and evaluation. It used public datasets and consumer-grade hardware to make replication easy.
But the potential is obvious. A system that can identify people using Wi-Fi alone, without light, without line of sight, and without physical contact, could find a place in future surveillance tools.
For now, it's a study. But the building blocks are in place.
Read next: Google Search Adds Custom Price Alerts and Virtual Clothing Try-On
