WiFi CSI: Your Router Can See You Move — Privacy Implications of Channel State Information

By Ken Imoto — 8-year Software Engineer at Propel-Lab, Author of "Practical Claude Code" on Kindle

What if I told you your home WiFi router could be acting as a surveillance camera right now?

In 2022, Carnegie Mellon University published "DensePose from WiFi," a paper demonstrating that WiFi signals alone can reconstruct 24-point human body poses through walls in real-time — with near-camera-level accuracy in controlled conditions. In February 2026, the implementation repository wifi-densepose hit GitHub Trending #1, reigniting attention around this technology.

No cameras. No microphones. If you're within WiFi range, your body position can be seen through walls. And this attack can be executed with an ESP32-S3 (~$30 USD).

I have personal experience with this technology. Back in 2022, I read the original paper and spent weeks trying to build human presence detection using consumer WiFi hardware. I consulted with a reinforcement learning researcher and iterated on implementations, but ultimately hit a wall: without CSI-capable hardware, reliable detection was impossible. The frustration stuck with me.

That's why the current situation concerns me. The barrier that stopped me in 2022 — "you need specialized hardware" — can now be crossed for $30.

In this article, I analyze the security risks from an engineer's perspective and explore what privacy design needs to look like going forward.

What Is WiFi CSI (Channel State Information)?

Channel State Information (CSI) is physical-layer data that describes the state of WiFi radio waves in transit. Specifically, it contains the amplitude and phase information for each OFDM subcarrier.

In normal WiFi operation, CSI is used for channel estimation — optimizing communication quality. However, because human bodies reflect, absorb, and diffract radio waves, CSI inherently contains information about who is in a room and what posture they're in.

The Technical Pipeline

Step by step:

1. WiFi router transmits radio waves — Standard beacon frames (always broadcasting)
2. Receiver extracts CSI — Records amplitude and phase changes per subcarrier
3. Deep learning model infers — Generates UV coordinate maps from CSI data
4. DensePose output — Reconstructs 24-point human body positions

In CMU's paper, training data was collected by simultaneously capturing camera footage and CSI data. DensePose labels were generated from camera images, then a model was trained to estimate those labels from CSI alone. Once trained, the camera is no longer needed at inference time.

CSI-Capable Hardware

CSI extraction doesn't require exotic equipment:

Device	Price Range	Notes
ESP32-S3	~$15-30 (dev board)	CSI-capable. Cheap and widely available
Intel 5300 NIC	~$30-50	Widely used in research
Nexmon-compatible chips	~$50+	Works with Raspberry Pi

The bottom line: attack hardware cost is approximately $30.

Security Risk Analysis

The security risks of WiFi CSI pose estimation are fundamentally different from traditional surveillance methods.

Risk 1: Passive Eavesdropping

The most critical threat: CSI attacks are completely passive. The attacker only "listens" to radio waves — they transmit nothing.

• No way to detect the attacker's presence
• IDS/IPS cannot catch it — no network anomaly to detect
• No trace in network logs — the attacker never connects to your network

Traditional surveillance requires physical installation of cameras or microphones. CSI attacks can be executed from an adjacent room or outside the building.

Risk 2: Lifestyle Pattern Analysis and Stalking

24-point pose estimation means the following information is available through walls:

• Home/away detection — presence or absence of human bodies
• Sleep/wake schedule — lying down vs. standing up movements
• Occupant count — identifying multiple people
• Daily routine patterns — eating, bathing, exercising schedules
• Visitor frequency and duration — inferring social relationships

This data could be weaponized for stalking or burglary reconnaissance.

Risk 3: Institutional Surveillance

State and corporate abuse scenarios are realistic:

• Warrantless surveillance — Legal frameworks don't address passive WiFi signal interception for body tracking
• Retail behavior tracking — Analyzing customer movement in stores without consent
• Employee monitoring — Tracking office behavior patterns
• Political surveillance — Monitoring movement at gatherings and meetings

Why Existing Defenses Don't Work

Here's the critical point: WiFi encryption (including WPA3) is completely powerless against CSI attacks.

The reason is straightforward. CSI operates at the Physical Layer (Layer 1) of the OSI model. Encryption protocols like WPA3 protect payloads at the Data Link Layer (Layer 2) and above. The amplitude and phase patterns of the radio waves themselves remain unchanged regardless of encryption.

Encrypted WiFi traffic doesn't change the physics. Radio waves still pass through walls and reflect off human bodies. This is a physics problem, not a crypto problem.

Current Defenses and Their Limitations

Let's evaluate what actually works:

Effective Countermeasures

1. Turn WiFi Off Completely (Effectiveness: 100%)

No radio waves, no CSI attack. But considering IoT devices, smart home systems, and remote work — this is impractical for most people.

2. RF Shielding (Effectiveness: 90%+)

Using RF-shielding wallpaper or paint to prevent WiFi signal leakage outside your space. Technically effective, but expensive to install and impossible in rental properties.

3. CSI Randomization (Effectiveness: 60-80%, Research Stage)

Injecting random noise into CSI at the router firmware level. Same concept as MAC address randomization, applied to the physical layer:

• Theoretical impact on communication quality is limited
• Could be deployed via firmware updates
• Currently research-only — no commercial router implementations exist

import numpy as np

def randomize_csi_preamble(original_preamble, privacy_level=0.5):
    """
    Add random phase rotation to beacon frame preambles
    to obscure human body information in CSI.

    Args:
        original_preamble: Original CSI preamble data
        privacy_level: 0.0 (no privacy) to 1.0 (max privacy)
                       Higher values = more noise = better privacy
                       but potentially more communication impact
    """
    num_subcarriers = len(original_preamble)

    # Generate random phase shifts
    phase_noise = np.random.uniform(
        -np.pi * privacy_level,
        np.pi * privacy_level,
        num_subcarriers
    )

    # Apply phase rotation
    # Communication quality impact is bounded by privacy_level
    randomized = original_preamble * np.exp(1j * phase_noise)
    return randomized


# Example: Different privacy levels
original = np.ones(64) * np.exp(1j * np.pi / 4)  # 64 subcarriers

low_privacy  = randomize_csi_preamble(original, privacy_level=0.2)
high_privacy = randomize_csi_preamble(original, privacy_level=0.8)

print(f"Phase variance (low):  {np.var(np.angle(low_privacy)):.4f}")
print(f"Phase variance (high): {np.var(np.angle(high_privacy)):.4f}")

Ineffective Countermeasures

Let's be crystal clear about what doesn't work:

Defense	Why It Fails
WPA3 encryption	CSI is a physical-layer property; encryption doesn't affect it
VPN	Encrypts traffic payload, not radio wave physics
MAC address randomization	CSI doesn't depend on MAC addresses
Firewall	Operates at network layer and above; physical layer is out of scope
Hidden SSID	Beacon frames are still transmitted (just without SSID field)

Thinking About Privacy Design

Given the WiFi CSI threat, here are three directions worth considering as engineers:

1. Embedding CSI Privacy Protection in WiFi Standards

The next IEEE 802.11 revision (802.11bn / WiFi 8) could potentially include CSI randomization as an optional feature. Specifically, adding random phase rotations to beacon frame preambles would make it harder to extract body information from CSI while minimizing communication quality impact.

The tradeoff between privacy and signal quality would need careful calibration:

Privacy Level vs Communication Impact (Theoretical)

Privacy │
        │                          ╱
  High  │                       ╱
        │                    ╱
  Med   │                 ╱      ← Sweet spot?
        │              ╱           (70% privacy,
  Low   │           ╱               <5% throughput loss)
        │        ╱
  None  │─────╱────────────────────
        └──────────────────────────
        None  Low   Med   High
              Communication Impact

2. Legal and Regulatory Frameworks

Current wiretapping and unauthorized access laws were not designed for passive WiFi CSI body tracking. Future discussions might address:

• Radio wave regulation expansion — Classifying passive WiFi interception for body information as a regulated activity
• Privacy law interpretation — Explicitly defining posture/movement data as personal information
• Building code updates — Establishing RF shielding standards for multi-unit housing

3. What Engineers Can Do Today

Setting aside large-scale policy changes, individual actions matter:

# Reduce router transmit power (OpenWrt example)
# Limits the physical range where CSI attacks are viable
uci set wireless.radio0.txpower=10  # Default is often 20-23
uci commit wireless
wifi reload

# Prefer 5GHz band — lower wall penetration than 2.4GHz
# Configure your devices to connect to 5GHz networks when available

Additional practical steps:

• Minimize transmit power — Reduce your WiFi signal's reach beyond your walls
• Prefer 5GHz — Lower wall penetration compared to 2.4GHz
• Reduce beacon interval — Fewer beacons = less CSI data available (at the cost of device discovery speed)
• Raise awareness — Share this knowledge with your team and organization

Attack Surface Comparison

To put WiFi CSI in perspective against other surveillance methods:

WiFi CSI uniquely combines low cost, zero detectability, through-wall capability, and passive operation. No other surveillance technology hits all four.

The Bigger Picture: Physical-Layer Security

WiFi CSI is one instance of a broader category: physical-layer information leakage. Similar principles apply to:

• Electromagnetic emanations (TEMPEST) — Reconstructing screen contents from EM radiation
• Power analysis — Extracting cryptographic keys from power consumption patterns
• Acoustic side channels — Inferring keystrokes from typing sounds

The common thread: security models that only consider Layer 2+ are incomplete. As engineers, we need to expand our threat models to include the physical layer, especially as cheap hardware makes these attacks accessible.

Key Takeaways

1. WiFi CSI + deep learning reconstructs 24-point human body poses through walls with camera-level accuracy
2. Attack cost is ~$30 (ESP32-S3), completely passive, and undetectable
3. All WiFi encryption including WPA3 is powerless — this is a physical-layer problem, not a crypto problem
4. The only realistic defense is CSI randomization, which is still in the research stage
5. Both WiFi standard-level privacy design and legal frameworks need updating

The security assumptions we've relied on — "encrypt it and you're safe," "firewalls protect us" — don't apply when the attack operates below the encryption layer. This is worth fundamentally rethinking how we approach security.

References

• Geng, J., et al. "DensePose from WiFi." arXiv:2301.00250 (2022) — Paper
• GitHub: ruvnet/wifi-densepose — Implementation repository
• IEEE 802.11 Standards — Official
• Ma, Y., et al. "WiFi Sensing with Channel State Information: A Survey." ACM Computing Surveys (2019)
• TEMPEST: NSA/CSS EPL-listed equipment standards for EM emanation security

---

I'm Ken Imoto, Software Engineer (8 yrs) at Propel-Lab. Author of "LLMO" and "Practical Claude Code" on Kindle. I write about security, AI, and emerging privacy threats. Follow me for more deep dives into the intersection of security and emerging technology.