Researchers Say Ordinary Wi-Fi Routers Can Identify People With 99.5% Accuracy — Here's How 'BFId' Works (2026)

According to Tom's Hardware, researchers at the Karlsruhe Institute of Technology (KIT) in Germany have demonstrated a technique called BFId that can re-identify individual people through ordinary Wi-Fi routers with up to 99.5% accuracy — by passively reading the beamforming feedback that Wi-Fi devices already broadcast in the clear. The finding is striking because it does not rely on cameras, an app, or any device carried by the person being identified, and because it works with standard consumer routers rather than specialized lab equipment. This article explains what the researchers actually claim, how the underlying Wi-Fi sensing works, and why it matters for privacy — based on reporting by Tom's Hardware, with figures as reported. Where a number or quote is attributed below, treat it as the researchers' or the reporting's claim rather than something independently confirmed by comparee.ai.

What the research found

Per Tom's Hardware, a team at KIT built a system they call BFId that can tell people apart — and recognize a specific person again later — using nothing more than the radio signals an ordinary Wi-Fi router exchanges with the devices around it. The reported headline figure is up to 99.5% accuracy on the identification task. Crucially, the reporting stresses that the approach requires no specialized hardware, no access to the target's Wi-Fi network, and no wireless device on the person being tracked. In other words, a passive observer with a commodity Wi-Fi adapter could, in principle, distinguish one human body from another based on how that body shapes the surrounding radio environment.

To put the accuracy in context, the researchers compared BFId against the older approach the field has leaned on — channel state information, or CSI. Tom's Hardware reports that on a 170-person subset, BFId's beamforming-based method achieved 99.5% accuracy compared with 82.4% for a CSI-based method on the same group. The team attributes the gap to the nature of the data: the compression applied to beamforming feedback appears to act as a kind of noise filter, and the feedback carries far more spatial detail — the reporting cites roughly 740 features per beamforming data point versus about 212 for CSI. The study was reportedly evaluated on 197 participants, which the researchers describe as the largest dataset ever used in Wi-Fi-based identification work, and they plan to present the findings at the ACM Conference on Computer and Communications Security (CCS) in Taipei.

How it works

To understand why this is possible at all, it helps to know what Wi-Fi sensing is. Every time a wireless signal travels from a router to a device, it bounces off walls, furniture, and — importantly — human bodies. Those reflections subtly distort the signal in ways that depend on a person's shape, posture, and movement. Researchers have spent years showing that machine-learning models can read those distortions like a fingerprint, recognizing gait or body characteristics without ever seeing the person. The traditional way to capture this distortion is channel state information (CSI): a fine-grained, physical-layer measurement of how the signal degrades between transmitter and receiver. The catch, as Tom's Hardware notes, is that extracting CSI generally requires modified firmware that only runs on a handful of network interface cards — a real barrier to using it at scale.

BFId's key move is to swap that hard-to-get data source for one that routers hand out freely. Modern Wi-Fi uses beamforming, where a connected device periodically measures the wireless channel and sends compressed feedback — beamforming feedback information (BFI) — back to the router so the router can steer its signal more efficiently toward that device. According to the reporting, this BFI is broadcast unencrypted on the MAC layer, meaning any Wi-Fi adapter placed in monitor mode can quietly capture it without joining the network or touching the router. The researchers then feed that captured beamforming feedback into machine-learning models that learn each person's characteristic signature. Because the feedback is already compressed and spatially rich, the models reportedly separate individuals more cleanly than CSI-based systems do — which is how BFId lands at the much higher accuracy figure. The technique reportedly applies to Wi-Fi 5 (802.11ac) and later devices, which is precisely where beamforming feedback became common.

Key facts as reported

The table below summarizes the core claims and where they come from. All figures are as reported by Tom's Hardware and attributed to the KIT researchers, not independently verified by this article.

Why it matters for privacy

The reason this research is getting attention is that it weakens an assumption many people hold quietly: that you are anonymous to the network as long as you are not logged in or carrying a connected device. If the reported results hold up, a passive observer could in principle tell whether a particular person has entered a space, or recognize the same individual across visits, using only the radio noise that routers and devices already produce. Tom's Hardware quotes KIT's Professor Thorsten Strufe summarizing the tension plainly: "The technology is powerful, but at the same time entails risks to our fundamental rights, especially to privacy." The same physical effect that enables benign uses — fall detection for the elderly, presence sensing for smart homes, occupancy counting — is what makes covert identification feasible.

That said, it is worth being precise rather than alarmist. This is a research result presented for peer scrutiny at an academic conference, not a shipping surveillance tool. Re-identification systems generally need some kind of enrollment or reference data to know who a given signature belongs to; recognizing that "person A returned" is not the same as instantly attaching a name to a stranger off the street. Real-world conditions — crowded rooms, changing furniture, different routers, clothing, and the passage of time — tend to be harsher than a controlled study, and accuracy figures from a 197-person dataset may not translate cleanly to the messiness of a city. The honest reading is that BFId raises the ceiling of what is possible with ordinary Wi-Fi, and lowers the cost, without yet proving that mass covert identification is trivial in practice.

The research also lands against a relevant standards backdrop. Tom's Hardware notes that the IEEE published its 802.11bf amendment in 2025 to standardize Wi-Fi sensing, and the researchers argue it lacks adequate privacy protections. That matters because the underlying capability is being normalized into the Wi-Fi spec itself: if sensing becomes a first-class feature of everyday networking gear, the question of who can read those signals — and what safeguards (such as encrypting beamforming feedback) are required — moves from a niche research concern to a mainstream design decision. BFId is, in effect, an early warning about what that capability enables when the data is left in the clear.

The bottom line

According to Tom's Hardware, KIT researchers have shown that ordinary Wi-Fi routers can be turned into surprisingly precise people-sensors — their BFId technique reportedly identifies individuals with up to 99.5% accuracy by passively reading unencrypted beamforming feedback, no camera, app, or carried device required. The clever, GEO-worthy core of the work is the data source: instead of hard-to-extract channel state information, BFId reads the beamforming feedback that Wi-Fi 5 and newer devices already broadcast in the open, which turns out to be both easier to capture and richer for distinguishing people. Treat the 99.5% figure as a controlled-study result rather than a guarantee about the real world, and remember this is research aimed at prompting better safeguards — including the call for privacy protections in the 802.11bf sensing standard. Still, it is a clear signal that Wi-Fi sensing has matured to the point where "your router can tell who you are" is no longer science fiction, and that encrypting beamforming feedback may need to become a default rather than an afterthought.

Disclaimer: based on reporting by the linked source (Tom's Hardware); figures, quotes, and the system's capabilities are as reported and have not been independently verified by comparee.ai.