BFId: How Wi-Fi Beamforming Feedback Can Identify People Without Cameras or a Device

Researchers at the Karlsruhe Institute of Technology demonstrated BFId, a Wi-Fi sensing attack that can infer a person’s identity from beamforming feedback information, achieving 99.5% accuracy in a controlled study involving 197 participants. The finding highlights a growing privacy problem in modern Wi-Fi networks: wireless data intended to improve connectivity can also reveal information about the people moving through an environment.

A person does not necessarily need to be carrying a smartphone, smartwatch, laptop, or other Wi-Fi-enabled device for the technique to work. Instead, BFId exploits how people physically alter radio waves as they move through an environment and how modern Wi-Fi clients report information about the wireless channel to access points.

The research, conducted by Julian Todt, Felix Morsbach, and Thorsten Strufe of the Karlsruhe Institute of Technology’s KASTEL Security Research Labs, represents one of the clearest demonstrations yet that Wi-Fi infrastructure can provide enough information for machine-learning systems to distinguish individuals. The researchers presented BFId at the ACM SIGSAC Conference on Computer and Communications Security (CCS) 2025.

The important caveat is that this does not mean every Wi-Fi router can currently identify every person who walks past it. BFId is a research attack evaluated under controlled conditions with a trained machine-learning model and a defined population of subjects. Nevertheless, the result demonstrates that the privacy signal exists and can be extracted from wireless infrastructure that was never designed to function as a biometric sensor.

What Is BFId?

BFId stands for an identity-inference attack using Beamforming Feedback Information (BFI).

Beamforming is a wireless transmission technique designed to make Wi-Fi communication more efficient. Modern Wi-Fi access points and client devices can use multiple antennas to manipulate the phase and direction of radio transmissions, effectively concentrating energy toward a receiver instead of transmitting it uniformly in every direction.

The technique became particularly important with Wi-Fi 5 (IEEE 802.11ac) and continues into newer generations of Wi-Fi. To perform beamforming effectively, devices need information about the wireless channel between the transmitter and receiver.

That feedback contains information about how radio waves are propagating through the environment. And that is where the privacy problem begins.

A Wi-Fi signal does not travel through an empty mathematical space. It interacts with walls, furniture, doors and, importantly, human bodies. People absorb, reflect, scatter and otherwise alter electromagnetic waves as they move through a room.

Consequently, measurements associated with the wireless channel can contain information about what is happening in the physical environment.

The KIT researchers demonstrated that BFI can contain enough information about those changes to support identity inference, rather than merely detecting whether somebody is present. Their study recorded BFI and Channel State Information from 197 participants and found that the BFI-based approach could recognize individuals with an accuracy of 99.5% ± 0.38 percentage points in their evaluation.

Why This Is Different From Conventional Wi-Fi Tracking

Traditional Wi-Fi tracking generally relies on identifying a device.

For example, a network operator might observe a device’s MAC address, IP address, authentication events, association history or other network metadata. Modern privacy protections have also made some forms of device tracking more difficult, including mechanisms such as randomized MAC addresses.

BFId approaches the problem differently. It does not fundamentally need to identify the device being carried by the person. Instead, it attempts to identify the physical person through the way that person’s presence and movement affect wireless signals.

If a person leaves their smartphone at home, device-based tracking obviously loses one source of information. But if the sensing system is observing how bodies perturb radio waves generated by other devices in the environment, leaving the phone behind does not necessarily eliminate the physical signal.

KIT explicitly notes that the person being observed does not need to carry a Wi-Fi device. The researchers argue that other active Wi-Fi devices in the surrounding environment can provide the wireless activity from which the sensing information is obtained.

How Wi-Fi Can “See” a Person

The underlying physics is not new. Radio waves interact with objects as they propagate. If something moves through the path of a wireless signal, the received signal can change.

Imagine a Wi-Fi access point transmitting toward a client. With an empty room, the signal follows one set of propagation paths. Place a person between or near the antennas and some of those paths change. Move the person and the propagation characteristics change again.

The resulting signal is not simply “strong” or “weak.” Depending on the system and measurement, changes can appear in properties such as amplitude, phase and spatial relationships across antennas and subcarriers.

This is one reason Wi-Fi sensing has become an active research area. Researchers have previously demonstrated that Wi-Fi signals can be used for applications ranging from occupancy detection and activity recognition to localization and, under specialized conditions, human pose estimation. NIST has described IEEE 802.11bf as an effort to enhance WLAN sensing for applications including presence detection and environmental monitoring.

BFId takes this concept into a more privacy-sensitive direction.

Instead of asking:

“Is there a person here?”

the researchers ask:

“Which person is this?”

That shift from presence sensing to identity inference is what makes the result particularly significant.

What Is Beamforming Feedback Information?

To understand BFId, it helps to distinguish BFI from the more commonly discussed Channel State Information (CSI).

CSI describes characteristics of a wireless channel as observed by a receiver. In simplified terms, it captures how a transmitted signal has been altered by the propagation environment.

Wi-Fi beamforming uses related channel information to calculate how transmissions should be spatially directed. In a typical beamforming exchange, a client provides the access point with feedback representing characteristics of the wireless channel. The feedback is a compressed representation rather than a complete raw channel measurement.

The BFId researchers found that this compressed representation still preserves information that can be useful for sensing. Earlier research had already demonstrated that beamforming feedback could be extracted from commercial Wi-Fi hardware and used for sensing applications. A 2023 study, for example, introduced Wi-BFI, a system for extracting beamforming feedback angles and reconstructing BFI from captured Wi-Fi frames on commercial IEEE 802.11ac and 802.11ax networks.

The BFId work builds on this broader research direction but focuses specifically on the privacy implications of using BFI to identify people.

Why Encryption of Normal Wi-Fi Traffic Does Not Automatically Solve the Problem

One of the most important technical details is that BFI occupies a different layer of the wireless communication process from the encrypted application traffic users normally think about.

Your HTTPS session may be encrypted. Your Wi-Fi data payload may be protected by WPA2 or WPA3. That does not automatically mean every piece of physical-layer information exchanged by Wi-Fi equipment is encrypted in the same way.

The BFId paper specifically investigates BFI because the relevant feedback information is transmitted in a form that can be observed over the air. Previous research likewise describes beamforming feedback angles as being transmitted in clear text and demonstrates that BFI can be reconstructed from captured frames.

This creates an unusual security situation. The attacker is not necessarily trying to decrypt somebody’s web traffic. Instead, the attacker is extracting physical information about the wireless channel itself. That means conventional advice such as “use encrypted Wi-Fi” does not completely address the sensing problem.

The BFId Experiment: 197 People

The scale of the KIT experiment is one of the most important reasons the research deserves attention.

The researchers created a dataset containing Wi-Fi recordings from 197 individuals, including BFI and CSI measurements. KIT describes it as its largest Wi-Fi sensing dataset containing both BFI and CSI recordings for this research.

Participants were recorded under different conditions, including different perspectives and walking styles. The researchers then trained machine-learning models on the resulting wireless time-series data. The BFI-based attack achieved:

99.5% ± 0.38% identification accuracy.

The paper also reports that BFI-based identification remained effective across different walking styles and perspectives and that the BFI approach performed better than comparable CSI-based approaches as the population increased. If a system recognizes two people in a laboratory, that is interesting.

Demonstrating strong performance across nearly 200 individuals is considerably more significant because the classification problem becomes harder as the number of possible identities increases.

Does the Person Need a Smartphone?

No, not according to the attack model demonstrated by the researchers. This is arguably the most counterintuitive aspect of BFId. The sensing system is not necessarily looking for a radio transmission originating from the target’s smartphone. Instead, it observes wireless channel information produced by communication involving other Wi-Fi devices.

When the target walks through the environment, their body changes the propagation characteristics of those radio signals. The machine-learning model can then learn patterns associated with those changes.

In practical terms, that means:

• A person does not necessarily need to be connected to the Wi-Fi network.
• The person does not necessarily need to carry a Wi-Fi device.
• Turning off a smartphone’s Wi-Fi does not necessarily eliminate the sensing signal.
• Other active Wi-Fi clients can provide the wireless activity needed for observation.
• The sensing system can potentially operate without a conventional camera.

KIT researchers specifically highlight this property as one of the privacy risks associated with the technique.

Does This Mean a Router Can Identify Anyone Who Walks Past It?

No. This is where headlines can become misleading. The 99.5% figure should not be interpreted as “every router can identify every stranger with 99.5% accuracy.”

BFId was evaluated using a controlled dataset containing known participants. The machine-learning system was trained to recognize individuals represented in that dataset.

This is essentially a closed-set identification problem.

If the model has learned the wireless signature associated with Alice, Bob and Charlie, it can classify a new observation against those known identities.

That is different from encountering an unknown person in a shopping mall and somehow discovering their name from nothing.

The researchers themselves frame the work as demonstrating an identity-inference threat rather than claiming that arbitrary Wi-Fi networks can universally determine the identity of every unknown person.

Why Gait Matters

One reason wireless identity inference is possible is that human movement is not perfectly random. People have characteristic walking patterns. Stride length, timing, body movement, arm motion, posture and other physical characteristics can influence how a person’s body interacts with radio waves.

A camera can capture these characteristics visually. BFId attempts to capture their radio-frequency consequences. The system does not need to produce a conventional photograph of a person’s face. Instead, machine learning can operate directly on time-series information derived from the wireless channel.

This is one reason the system can continue to distinguish people across changes in viewing perspective and walking style. The important idea is not that Wi-Fi magically produces a normal camera image.

Rather, the wireless signal contains structured information about the physical scene, and machine-learning models can learn patterns within that information.

BFI vs. CSI: Why the New Attack Matters

Wi-Fi sensing research has traditionally paid considerable attention to CSI.

CSI can provide detailed information about how wireless signals change across frequencies and antenna paths, making it useful for sensing human activity.

But CSI-based sensing can impose stronger requirements depending on the implementation.

BFId is particularly interesting because BFI provides a different observation point. The KIT researchers say their BFI-based attack requires a weaker adversary model than earlier CSI approaches. Their experiments found that BFId could achieve high identification accuracy without the specialized hardware and custom firmware associated with some previous approaches.

That matters from a security perspective. A privacy attack becomes more concerning when the barrier to collecting the underlying data becomes lower.

The research therefore shifts the question from:

“Can sophisticated Wi-Fi sensing identify people?”

to:

“Can information already exchanged by ordinary Wi-Fi equipment be repurposed for identity inference?”

The answer demonstrated by BFId is yes.

The Surveillance Implications Are Bigger Than a Home Router

The privacy implications extend beyond individual homes.

Wi-Fi infrastructure is now present across offices, universities, hotels, restaurants, airports, retail environments and public spaces.

A conventional camera is visible.

It has a physical location.

In many jurisdictions, camera surveillance may also be subject to signage, disclosure or other legal requirements.

Radio-based sensing can be considerably less obvious.

A person walking through an environment may not realize that their movements are influencing wireless measurements.

KIT researchers specifically warned that ubiquitous Wi-Fi infrastructure could potentially become a surveillance layer. They highlighted scenarios in which someone repeatedly passing a Wi-Fi-equipped location could potentially be recognized later.

CCTV observes reflected light. BFId demonstrates that wireless infrastructure can observe the physical consequences of reflected and propagated radio waves.

The technology is not inherently malicious. Wi-Fi sensing itself has legitimate applications.

But the same capability can be repurposed.

Wi-Fi Sensing Is Not Automatically a Security Vulnerability

BFId is better understood as a privacy attack or side-channel risk than a conventional router software vulnerability such as a memory corruption flaw or remote-code-execution bug.

There is not necessarily a defective router implementation that can simply be patched with a conventional security update. The underlying issue is that wireless protocols expose information required for legitimate communication, and that information can contain additional information about the physical environment.

This makes mitigation significantly harder.

It is similar to other side-channel problems where information that is harmless for one purpose becomes sensitive when combined with machine learning or another analysis technique.

The Connection to IEEE 802.11bf

The research arrives as Wi-Fi sensing itself becomes increasingly standardized.

IEEE 802.11bf is specifically focused on enhancing WLAN sensing capabilities. NIST describes the work as supporting sensing applications involving presence detection, environment monitoring and other measurements, while the IEEE describes the completed amendment as IEEE 802.11bf-2025, covering enhancements to WLAN sensing operation across specified license-exempt frequency bands.

This is an important correction to some earlier descriptions of the research that refer to 802.11bf as a merely “forthcoming” standard.

As of 2026, IEEE 802.11bf-2025 is an active standard.

That does not make the BFId research obsolete. If anything, it makes the privacy discussion more relevant. The industry is actively expanding the ability of Wi-Fi systems to sense their surroundings.

That creates legitimate opportunities for applications such as smart-building automation, occupancy detection, localization and other sensing systems. But it also means privacy safeguards need to be considered alongside sensing capabilities.

Why 802.11bf Creates a Broader Privacy Conversation

IEEE 802.11bf formalizes mechanisms for WLAN sensing, including measurements that can be used to estimate characteristics such as range, velocity, angle and motion. The IEEE’s WLAN sensing work explicitly considers objects including humans and environments such as homes and enterprise spaces.

This creates a fundamental policy question:

Who owns the information generated when wireless infrastructure senses the physical environment?

A person may never connect to the network. They may never authenticate. They may never accept a privacy policy. Yet their physical presence can influence wireless measurements.

That is fundamentally different from conventional network telemetry associated with a subscriber or device. It is an emerging category of privacy-sensitive information generated indirectly from the environment.

Could Wi-Fi Become a Biometric Sensor?

BFId does not establish that BFI is legally a biometric identifier, and it would be inappropriate to automatically classify the resulting data as biometric information under every privacy regime.

Technically, however, the research demonstrates something that is biometric-like in behavior.

The system learns a physical pattern associated with a particular person and uses that pattern to distinguish them from other individuals.

The underlying signal is not a fingerprint scan or facial image.

It is a radio-frequency representation of how an individual’s physical characteristics and movement affect the wireless channel.

What Makes the 99.5% Result So Significant?

Accuracy alone does not tell the entire story.

The researchers evaluated the system using a relatively large population of 197 participants, and the reported accuracy was 99.5% ± 0.38%. They also tested different perspectives, walking styles and sample sizes.

The results suggest that the relevant wireless patterns are not simply a fragile artifact of one exact camera angle or one carefully repeated movement.

The system was designed around recurrent neural networks, processing temporal patterns in the wireless data rather than relying on a conventional visual image.

That is significant because human movement is inherently temporal. A single radio measurement may reveal little. A sequence of measurements can reveal a pattern. Machine learning can then exploit that pattern.

What an Attacker Would Actually Need

It is tempting to imagine an attacker simply pointing a router at someone and receiving a name.

That is not how the demonstrated system works. A realistic identity-inference deployment would require several components:

Wireless observations. The attacker needs access to relevant BFI measurements generated by compatible Wi-Fi activity.

Signal processing. Raw wireless observations need to be transformed into data that a machine-learning system can consume.

Training data. The system needs examples associated with the people it is expected to recognize.

A trained classifier. BFId uses machine learning to distinguish the resulting temporal patterns.

An appropriate observation environment. Radio propagation depends heavily on the physical environment, device configuration, traffic patterns and geometry.

This is why the research should not be described as a magical “Wi-Fi facial recognition” system. It is a sophisticated sensing pipeline.

The important security finding is that the pipeline can work using information already present in modern Wi-Fi communication.

What Happens if You Turn Off Wi-Fi on Your Phone?

Turning off Wi-Fi on your own phone can prevent your phone from generating or receiving certain wireless traffic. It does not, by itself, guarantee that your body cannot influence Wi-Fi sensing measurements. If other devices are communicating nearby, their radio waves can still interact with you.

That is the key distinction between device tracking and environmental sensing.

Device tracking asks:

“Where is this device?”

Wi-Fi sensing asks:

“What is happening to the radio environment?”

BFId demonstrates that the second question can potentially provide enough information to answer a third:

“Which known person is producing this pattern?”

Can You Protect Yourself From BFId?

For ordinary users, there is currently no simple universal privacy switch that guarantees protection from this class of sensing.

The problem is partly architectural.

If a Wi-Fi network relies on beamforming feedback for performance, eliminating or substantially changing that feedback may affect wireless functionality.

The most effective mitigation therefore needs to happen at the protocol and implementation level.

Potential approaches include:

Encrypting Sensitive Feedback

One obvious direction is protecting beamforming feedback so that passive observers cannot freely extract it.

However, encryption introduces protocol and interoperability considerations because the access point needs the information to perform its wireless function.

Reducing Feedback Exposure

Another potential strategy is reducing how frequently sensitive information is transmitted or limiting the precision of the feedback.

But this creates a tradeoff between privacy and wireless performance.

The BFId research is particularly important because it demonstrates that simply assuming compressed feedback is harmless is not a safe privacy model.

Privacy-Preserving Wi-Fi Sensing

Future standards and implementations could distinguish between sensing information needed for legitimate network operation and information that can expose sensitive characteristics of people in the environment.

This could involve access controls, privacy modes, constrained reporting or other protocol-level mechanisms.

Disabling Beamforming Where Appropriate

Some networking equipment provides configuration options related to beamforming.

Where such controls exist, disabling the feature could reduce exposure to attacks that specifically depend on beamforming feedback.

However, this should not be treated as a universal solution. Router capabilities differ, and disabling beamforming can affect performance.

More importantly, Wi-Fi sensing and wireless-channel privacy involve more than a single feature.

What Network Administrators Should Consider

Organizations operating Wi-Fi networks should treat wireless sensing as a privacy issue rather than solely a performance feature.

Enterprise environments should consider:

• Which Wi-Fi generations and sensing capabilities are deployed.
• Whether beamforming features can be configured or restricted.
• What wireless information is observable over the air.
• Whether wireless sensing is explicitly required for the business use case.
• Whether employees, visitors or customers could reasonably expect physical privacy.
• Whether future Wi-Fi sensing capabilities create additional privacy obligations.
• Whether security and privacy teams are evaluating PHY-layer information alongside conventional network logs.

This is especially relevant in environments where people may reasonably expect privacy, such as healthcare facilities, workplaces, residential buildings and sensitive government environments.

Why Conventional Privacy Controls May Not Be Enough

Many privacy controls focus on application-layer data.

They regulate things such as:

• Names
• Email addresses
• IP addresses
• Cookies
• Device identifiers
• Location data
• Browsing history

BFId highlights another category:

physical information inferred from wireless signals.

This information may never appear in an application’s database.

It can exist as radio measurements.

Machine-learning systems can then transform those measurements into higher-level conclusions.

That creates a growing challenge for privacy engineering.

The raw data may look meaningless to a human analyst.

The inference derived from it may be highly sensitive.

The Bigger Lesson for Cybersecurity

BFId is a reminder that security cannot stop at the application layer.

Modern wireless systems are enormously complex. They expose information across multiple layers because those layers need to cooperate to deliver reliable communication.

Every exposed signal can potentially become a measurement source. Every measurement source can potentially become a sensing channel. And machine learning makes it increasingly practical to extract information from signals that engineers previously considered too noisy, compressed or indirect to reveal sensitive information.

This is why physical-layer privacy deserves more attention. The next generation of security research will increasingly have to consider not just:

“Can an attacker read the data?”

but also:

“What can an attacker infer without reading the data?”

What BFId Does and Does Not Prove

The research establishes a strong technical result, but it is important not to overstate it.

BFId demonstrates:

• Identity inference from Wi-Fi beamforming feedback information.
• 99.5% ± 0.38% accuracy in the reported evaluation.
• Evaluation using 197 participants.
• Robustness across multiple perspectives.
• Robustness across different walking styles.
• A sensing approach that does not require the target to carry a Wi-Fi device.
• A privacy threat arising from information exposed through ordinary Wi-Fi operation.

BFId does not demonstrate:

• That every Wi-Fi router can identify every person.
• That an unknown stranger can automatically be named without prior training data.
• That Wi-Fi can literally produce a conventional photographic image of a person.
• That WPA2 or WPA3 encryption can simply be “cracked” through this technique.
• That every Wi-Fi environment will produce the same 99.5% accuracy.
• That turning off a person’s phone is the only privacy measure available.

Why This Research Matters Now

Wi-Fi has evolved from a basic connectivity technology into an increasingly capable sensing platform. The same radio infrastructure that delivers internet access can potentially detect movement, estimate environmental properties and infer characteristics of people.

IEEE 802.11bf-2025 formalizes the industry’s move toward more capable WLAN sensing. At the same time, BFId demonstrates that privacy risks can emerge from existing mechanisms before users ever interact with an explicitly marketed sensing feature.

That combination deserves attention from network engineers, security researchers, hardware vendors, regulators and privacy professionals.

The concern is not that Wi-Fi has suddenly become a camera. The concern is more subtle and arguably more important:

Wi-Fi does not need to behave like a camera to reveal information about the people around it.

Final Takeaway

The BFId research from Karlsruhe Institute of Technology demonstrates that beamforming feedback information, originally exchanged to help modern Wi-Fi systems optimize wireless communication, can also act as a powerful sensing signal.

Using recordings from 197 participants, the researchers reported 99.5% ± 0.38% identification accuracy, with the attack remaining effective across different perspectives and walking styles.

The most important privacy implication is that the target does not necessarily need to carry a connected device. The system can instead exploit the way a person’s body and movement alter radio propagation involving other active Wi-Fi devices.

That makes BFId fundamentally different from traditional device tracking.

It also demonstrates why future wireless security cannot be evaluated solely by asking whether communication is encrypted. A wireless protocol can protect the contents of a message while still exposing physical information that enables powerful inferences about the environment.

Wi-Fi sensing has legitimate and potentially valuable applications, but its privacy implications need to be addressed at the same time that the technology becomes more capable.

IEEE 802.11bf-2025 shows that WLAN sensing is moving from research toward standardized wireless functionality. The BFId research provides an important warning about what that future could look like if privacy is treated as an afterthought.

The biggest lesson is simple: you may not need to connect to a Wi-Fi network, carry a smartphone, or stand in front of a camera for wireless infrastructure to learn something about you. Your physical presence can be enough to change the signal.

FAQs

Research and Source Notes

The primary source for this report is the BFId research by Julian Todt, Felix Morsbach and Thorsten Strufe of Karlsruhe Institute of Technology’s KASTEL Security Research Labs, published as part of ACM CCS 2025. KIT’s publication record identifies the work as BFId: Identity Inference Attacks Utilizing Beamforming Feedback Information.

KIT has also published the BFId dataset containing BFI and CSI recordings from 197 individuals for non-commercial research access.

Independent background research on extracting BFI from commercial Wi-Fi hardware was published by Haque, Meneghello and Restuccia in 2023, providing additional technical context for why beamforming feedback can be captured and analyzed.

For the standards context, IEEE lists IEEE 802.11bf-2025 as an active amendment for WLAN sensing, while NIST has documented the technical direction and sensing capabilities associated with the standardization effort.

This post first appeared at - The CyberSec Guru