Wi‑Fi Beamforming Feedback Enables 99.5 % Person Identification – Security Implications for Everyday Routers

Announcement

Security researchers from the Karlsruhe Institute of Technology (KIT) have demonstrated a passive attack that can recognize a person walking through a room with 99.5 % accuracy using only the beamforming feedback information (BFI) emitted by ordinary Wi‑Fi routers. The technique, named BFId, requires no specialized firmware, no connection to the target network, and works even when the subject carries no Wi‑Fi device. The results, based on a dataset of 197 participants, will be presented at the ACM CCS conference in Taipei.

Technical specs

Beamforming feedback vs. channel state information

• BFI is a compressed representation of the channel matrix that a client sends to an access point (AP) during the beamforming handshake defined in 802.11ac (Wi‑Fi 5). The AP rebroadcasts this feedback on the MAC layer unencrypted, allowing any monitor‑mode adapter to capture it.
• CSI (Channel State Information) is a raw, high‑resolution measurement of the same channel. Extracting CSI historically required custom drivers (e.g., Intel 5300) and is supported by fewer than 6 % of deployed devices as of 2023.

Data characteristics

Metric	BFI	CSI
Features per sample	740	212
Compression ratio	~4:1 (lossy)	none
Capture scope	One packet contains feedback from all associated clients	One packet per client, per NIC
Required hardware	Any commercial Wi‑Fi 5/6 router + standard monitor‑mode NIC	Modified NIC firmware (Intel 5300 or similar)

Accuracy results

• BFI: 99.5 % identification on a 170‑person test subset.
• CSI: 82.4 % on the same subset. The authors attribute BFI’s superiority to two factors:

1. Implicit noise filtering – the compression discards high‑frequency variations that otherwise confuse classifiers.
2. Higher spatial diversity – each BFI packet aggregates measurements from multiple antennas, providing richer geometric cues.

Attack workflow

1. Passive sniffing – an attacker places a single Wi‑Fi adapter in monitor mode near the target AP.
2. Capture BFI frames – the adapter records every beamforming feedback frame broadcast by the AP.
3. Feature extraction – raw BFI payloads are parsed into a 740‑dimensional vector per frame.
4. Model inference – a pre‑trained machine‑learning model (e.g., gradient‑boosted trees) maps feature vectors to a user identity.
5. Real‑time tracking – because BFI is emitted every 10–20 ms per client, the system can update its prediction continuously as a person moves.

Mitigation attempts

Mitigation	Effect on accuracy
Reduce BFI reporting frequency (down to 1 Hz)	< 2 % drop – still > 95 %
Randomly drop BFI frames	Minor impact; classifier tolerates gaps
Encrypt BFI payloads	Would require a new amendment to the 802.11 standard; current hardware would become incompatible

The paper notes that the upcoming 802.11bf amendment (standardized in 2025) formalizes Wi‑Fi sensing but lacks explicit privacy safeguards, leaving the door open for the type of exploitation demonstrated by BFId.

Market implications

1. Immediate exposure for existing deployments – Home routers, enterprise APs, and IoT gateways that support 802.11ac/ax already broadcast BFI. Organizations that rely on Wi‑Fi for security monitoring (e.g., occupancy‑based HVAC control) must reassess the privacy risk.
2. Supply‑chain pressure – Chip manufacturers such as Broadcom, Qualcomm, and MediaTek, which provide the baseband processors for most routers, may see OEMs request firmware that either encrypts BFI or disables it by default. A firmware update cycle could become a competitive differentiator for security‑focused brands.
3. Regulatory response – European data‑protection authorities are likely to classify unencrypted BFI as personal data, triggering GDPR obligations for any entity that captures or processes it. Similar scrutiny could appear in the U.S. under state‑level privacy statutes.
4. Standard‑body activity – The IEEE 802.11 Working Group is expected to draft amendments that either (a) mandate optional encryption of BFI, or (b) introduce a “privacy‑mode” flag that suppresses broadcast of raw feedback. Vendors that adopt these changes early could market “privacy‑by‑design” routers.
5. Business‑model shift for security firms – Companies offering Wi‑Fi‑based presence detection (e.g., smart‑building platforms) may need to transition to edge‑only processing, where raw BFI never leaves the device, or to alternative sensors (ultrasonic, LiDAR) to avoid legal exposure.

Bottom line

The BFId study proves that standard, unencrypted beamforming feedback is a high‑fidelity biometric signal. With a single passive listener, an adversary can achieve near‑perfect person identification without any cooperation from the target. The findings pressure router manufacturers, chipset vendors, and standards bodies to embed privacy controls into the next generation of Wi‑Fi firmware, and they give regulators a concrete technical baseline for drafting privacy‑focused rules.

---

For the full research paper, see the pre‑print on arXiv.