INDIGO AirGuard
EUDIS 2026
Home Demo Paper Doctrine Technology Tour

Deep Technology

From physics intuition to distributed detection — the science behind turning 1.8 million smartphones into a city-scale early warning system.

∴

Detection Theory — The Core Innovation

01 / 07

From Individually Useless to Collectively Certain

A single smartphone at 500 m from a Shahed-136 has an acoustic SNR of approximately 0 dB. At this SNR, the detection probability per single observation is 1.8% — functionally useless. But AirGuard does not rely on a single phone.

1.8%
Single phone Pd at SNR = 0 dB
90,000
Active sensors (5% of 1.8M phones)
>99.97%
System Pd after cascade fusion
10−6
False alarm rate per hour per km²

The Neyman-Pearson Framework

Each phone faces a binary hypothesis test: is a drone present (H1) or not (H0)?

Likelihood Ratio Test Λ(x) = p(x | H1) / p(x | H0) Decide H1 if Λ(x) > γ Decide H0 otherwise Neyman & Pearson (1933), "On the Problem of the Most Efficient Tests of Statistical Hypotheses"

In practice, the YAMNet classifier output acts as a learned sufficient statistic that approximates the Generalized Likelihood Ratio Test (GLRT).

Non-Coherent Integration: The ^n Power

Multiple independent observations combine to improve the effective SNR linearly:

Energy Integration SNReff = N × SNRsingle For N = 1,000 phones at SNR = 0 dB: SNReff = 10 × log10(1000) + 0 = 30 dB Pd > 0.9999 at Pfa = 10−6 Marcum (1960), "A Statistical Theory of Target Detection by Pulsed Radar"

4-Level CFAR Cascade

Cell-Averaging CFAR adapts each phone's threshold to its local noise environment, maintaining a constant false alarm rate. The cascade amplifies true detections while self-limiting false alarms:

LEVEL 0
Phone
CA-CFAR per device
▶
LEVEL 1
Cell
K-of-N (500m×500m)
▶
LEVEL 2
Cluster
Adjacent cell correlation
▶
LEVEL 3
City
Track init + EKF

The Avalanche Algorithm

A detection simultaneously serves as an observation AND a command — it tells neighboring phones to listen harder. This "detection-is-command" protocol uses confidence-dependent alert radii with Gaussian spatial decay and temporal quench mechanisms.

Sensitization Protocol γj,new = γj × (1 − η × ci × exp(−dij² / 2Ralert²)) Alert radius: c ∈ [0.3, 0.5) → 200 m c ∈ [0.5, 0.7) → 500 m c ∈ [0.7, 0.9) → 1000 m c ∈ [0.9, 1.0] → 2000 m Quench: spatial coherence + confuser veto + rate limiting Novel architecture. Building blocks: Finn & Johnson (1968) CFAR, Varshney (1997) distributed detection, Wald (1945) SPRT.
A single phone has a 1.8% chance of detecting a Shahed at 500 m. When 20 phones in a city block share observations through the cascade algorithm, the combined probability exceeds 99.97%. Noise becomes signal through mathematical cooperation.
≡

The Equation of State

02 / 07

Thermodynamic Analogy for Detection

Just as the ideal gas law PV = nRT constrains thermodynamic systems, a fundamental constraint equation governs distributed detection systems:

AirGuard Equation of State Pd × Vsearch = k × Nsensors × Tdwell / Pfa−1
Pd
Detection probability
Vsearch
Surveilled volume
Nsensors
Active sensor count
Tdwell
Integration time
Pfa
False alarm rate

The trade-off: to cover more volume at the same Pd, you need either more sensors, longer dwell time, or you must accept a higher false alarm rate. AirGuard breaks this constraint by having 90,000 sensors already deployed — the "N" term is three orders of magnitude larger than any purpose-built system.

Bucharest Numbers

Coverage Calculation Volume: 600 km² × 3 km altitude = 1,800 km³ Cells: 36,000 resolution cells (500m × 500m × 200m) Active: 90,000 phones (5% of 1.8M) Density: 2.5 phones per cell (average) At 2.5 phones/cell: SNReff = 4 dB — insufficient alone. This is why the avalanche cascade is essential: it recruits neighboring cells, effectively multiplying N per region of interest.

Belief Dynamics: Stochastic Differential Equation

Detection confidence evolves in continuous time, connecting to established Kalman filtering theory:

Belief SDE dB(t) = [−λdecay × B(t) + Σk message_ratek + Σj ∈ neighbors μj × Bj(t − δj) + u(t)] dt + σobs dW(t) Partially novel formulation. Related: IPDA filters (Bar-Shalom), Kalman (1960). SDE formulation for smartphone mesh is original.
⌘

8 Novel Sensor Modalities

03 / 07

Eight sensing concepts evaluated with quantitative physics calculations, grounded detection ranges, and cost estimates. Tiered by readiness and impact.

Tier 1 — Build Now

Yagi + Servo PBR

Score: 8/10
Range: 1–2 km
Cost: 53 EUR/unit
Published DVB-T results. Cued by acoustic detection via NATS.
Tier 1 — Build Now

Noise2Weight

Score: 8/10
Range: 150–500 m
Cost: 0 EUR (software)
98% payload accuracy from 0.25s audio. Extends YAMNet, zero hardware.
Tier 1 — Build Now

WiFi CSI Mesh

Score: 7/10
Range: 77–195 m
Cost: 300 EUR total
Propeller blade matches 2.4 GHz wavelength — strong micro-Doppler.
Tier 2 — Post-Hackathon

FSR "Radio Blade"

Score: 9/10 physics
Range: 1–10 km baseline
Cost: ~5K EUR pilot
34 dB enhancement. Total stealth immunity. Flagship finding.
Tier 2 — Post-Hackathon

Geophone Array

Score: 7/10
Range: 200–2,000 m
Cost: 55–150 EUR/node
Wind-immune. Fills acoustic gap in bad weather.
Tier 3 — Niche

PM2.5 Chemical Plume

Score: 4/10
Range: 30–100 m
Cost: 60–80 EUR/node
Combustion confirmation only. Low-altitude niche.
Tier 3 — Research

LIDAR Wake

Score: 3/10
Range: 50–200 m
Cost: 100K+ EUR
Signal buried in atmospheric turbulence noise.
Tier 4 — Impossible

Gravity Gradient

Score: 0/10
Range: 0 m
Cost: 500K+ EUR
Signal 4–7 orders below sensor noise floor. Eliminated.
Research rigour: We evaluated 8 modalities including ones we could prove are impossible. The gravity gradient analysis (Tier 4) demonstrates that we do not just pursue ideas — we kill them when the physics does not support them.
⚡

The Forward-Scatter Breakthrough

04 / 07

The single most important finding from our sensor research: Forward-Scatter Radar provides 34 dB RCS enhancement over monostatic radar for Shahed-class targets.

Babinet's Principle: Stealth Is Irrelevant

σFS = (4π × A²) / λ²

In forward-scatter geometry, the target's radar cross section depends only on its physical silhouette — not on material properties, shaping, or coatings. A stealth-coated drone has the same FSR RCS as a chrome-plated one.

Shahed-136 (A ~ 0.5 m²) at 900 MHz cellular frequency:

σFS = 28.3 m² = +34.5 dB over monostatic (0.01 m²)

Uses existing cellular base stations as illuminators. No transmission license required. No new infrastructure needed.

Completely immune to all stealth technologies

Comparison: Monostatic vs. Forward-Scatter

Property Monostatic Radar Forward-Scatter Radar
Shahed-136 RCS 0.01 m² (−20 dBsm) 28.3 m² (+14.5 dBsm)
Stealth effectiveness Reduces RCS by 20–30 dB Zero effect (Babinet's principle)
Infrastructure Dedicated transmitter required Existing cellular/DVB-T towers
License required Yes (active RF emission) No (passive reception only)
RCS enhancement Baseline +34 dB
Forward-Scatter RCS — Derivation From Babinet's principle (optics diffraction theory): The scattered field from an opaque object equals the diffracted field from an aperture of the same shape. σFS = (4π A²) / λ² where A = physical cross-sectional area (silhouette) Quadcopter (A ~ 0.1 m²) at 500 MHz DVB-T: σFS = 4π × 0.01 / 0.36 = 0.35 m² (−4.6 dBsm) vs. monostatic: −30 to −20 dBsm Enhancement: 15–25 dB Abdullah et al. (2020), FSR DVB-S drone detection; Musa et al. (2019), micro-Doppler in FSR geometry.
≈

Interferometric Detection

05 / 07

Phase-Difference Angle-of-Arrival

Two or more receivers with known baseline d measure the phase difference of an incoming RF signal. The same principle that powers LIGO and radio astronomy VLBI, applied to drone tracking:

Interferometric AoA sin(θ) = (Δφ × λ) / (2π × d) Angular resolution: θres ≈ λ / d (radians) Phase interferometry is foundational. Applied to drone detection: Jian & Lu (2018), IEEE RadarConf.

Baseline Requirements by Frequency

Parameter FM (100 MHz) DVB-T (500 MHz) WiFi (2.4 GHz)
Wavelength (λ) 3.0 m 0.6 m 0.125 m
Baseline for 1° resolution 172 m 34 m 7.2 m
Baseline for 5° resolution 34 m 6.9 m 1.4 m
Half-wavelength (unambiguous) 1.5 m 0.3 m 6.25 cm

MUSIC/ESPRIT Super-Resolution

Multi-element arrays with MUSIC (MUltiple SIgnal Classification) or ESPRIT algorithms exploit the covariance matrix structure to estimate AoA below the Rayleigh limit. Published results: USRP-based 5-element array at 2.46 GHz achieved 3.27° mean angular error with MUSIC algorithm at ranges up to 2.5 km.

WiFi CSI Micro-Doppler

At 2.4 GHz, the propeller blade length (~12 cm) matches the wavelength — producing exceptionally strong micro-Doppler signatures. WiFi Channel State Information (CSI) captures these across 52–256 subcarriers at 100+ fps, enabling drone type classification at ranges where acoustic signatures fail.

Key insight: At 2.4 GHz, a drone propeller is an efficient antenna. Each blade rotation modulates the WiFi channel, creating a fingerprint as unique as an engine sound — but immune to wind noise.

Acoustic Interferometry

Applied to drone blade-pass frequencies (133–267 Hz for quadcopters): a medium-baseline acoustic interferometer (10–50 m) operating on BPF harmonics achieves 1–4° angular resolution — far superior to compact TDoA arrays.

Acoustic Interferometer Resolution 100 m baseline at 500 Hz (λ = 0.686 m): θres = 0.686 / 100 = 0.39° 1 km baseline at 500 Hz: θres = 0.039° GPS-PPS sync accuracy (<100 ns) is orders of magnitude sufficient for phase coherence at acoustic frequencies. Distributed arrays: PMC11946234 (2025), 0.77 m localization error. Fiber-optic DAS: NEC Labs (2022), 1.47° RMSE over 125 km.
★

What Is Genuinely Novel

06 / 07

Assessed against prior art by an independent innovation panel. We are honest about what is novel and what is not.

Novelty: 5/5

City-Scale Drone Detection from Civilian Smartphones

No prior system uses the existing installed base of civilian smartphones as a distributed acoustic sensor network for military-grade drone early warning. Every existing C-UAS system (DroneShield, Dedrone, Hensoldt, Sky Fortress) requires purpose-built hardware at EUR 50–100M per city. AirGuard: EUR 50–100K.

Closest analogy: MyShake (UC Berkeley, 2016) for earthquake detection — proves the distributed smartphone sensing model works, but does not cross into the military domain.
Novelty: 4/5

ACIR — Acoustic Channel Impulse Response for Drone Detection

Cross-domain transfer of the WiFi CSI detection paradigm to acoustics for airborne target detection. Conventional detection asks "can I hear the drone?" — ACIR asks "has anything changed in the acoustic environment?" A stealth drone too quiet to hear directly still shadows and diffracts ambient sound.

No published prior art found combining acoustic impulse response measurement with drone detection. Potential patent. TRL: proof of concept.
Novelty: 4/5

WiFi CSI Parabolic Dish + YAMNet Fusion

No one has combined parabolic-dish-focused WiFi CSI with acoustic YAMNet classification in a multi-modal pipeline. The propeller-wavelength match at 2.4 GHz produces uniquely strong micro-Doppler for drone type classification.

Genuinely novel combination — potential patent/publication. Each component exists; the fusion is original.
Novelty: 3/5

Distributed CFAR Avalanche Cascade

Individual building blocks are established (CFAR 1968, K-of-N 1997, SPRT 1945). The novel contribution: 4-level hierarchical cascade with "detection-is-command" sensitization, confidence-dependent alert radius, and quench mechanisms for correlated false alarms.

Engineering innovation — novel architecture from proven components. This is how successful defence systems are designed.
Novelty: 3/5

Belief SDE Framework

Continuous-time stochastic differential equation formulation for detection dynamics in smartphone mesh networks. Extends IPDA track existence probability into a full SDE with spatial coupling.

IPDA filters exist; this SDE formulation for smartphone sensing mesh is partially novel.
What we do NOT claim as novel: Acoustic drone detection (mature field), YAMNet classification (published since 2019), passive bistatic radar (Fraunhofer, NATO studies), EKF tracking (textbook), CoT/ATAK output (NATO standard). We use established techniques where they work and innovate only where the problem demands it.
△

8-Layer Detection Cascade

07 / 07

Each layer narrows uncertainty. An FSR crossing triggers a PBR scan, which cues WiFi CSI focus, which steers acoustic beam forming, which confirms via payload estimation, geophone confirmation, and chemical plume analysis. The avalanche cascade realized across 8 modalities.

0
FSR "Radio Blade" Fence
Outer perimeter, 1–10 km, stealth-immune forward scatter
Post-hackathon
1
Distributed Acoustic (YAMNet + CFAR)
All ranges to 500 m, ML classification, 90K smartphones
Exists
2
WiFi CSI Micro-Doppler
Close-range <200 m, propeller-wavelength match, type classification
Hackathon
3
Passive Bistatic Radar (DVB-T)
Mid-range 1–3 km, range-Doppler extraction, Yagi scanning
Hackathon
4
Acoustic Interferometry (TDoA)
GPS-PPS fixed nodes, sub-degree AoA, precision localization
Exists
5
Geophone Seismic Array
200–2,000 m, wind-immune, acoustic-seismic coupling
Post-hackathon
6
RF Direction Finding
Remote ID, control link DF (commercial drones only)
Exists
7
Chemical Plume (PM2.5)
<30 m altitude, combustion engine confirmation
Niche
Defence in depth: No single modality is sufficient. Each layer compensates for another's weakness. Acoustic fails in wind — geophones compensate. RF fails against INS/GLONASS drones (Shahed) — acoustic compensates. Stealth defeats monostatic radar — FSR is immune. The cascade is the innovation.

Key References

Detection theory: Neyman & Pearson (1933); Finn & Johnson (1968); Wald (1945); Marcum (1960); Varshney (1997).

Passive radar: NATO STO MSG-SET-183; Martelli et al. (2020); Chadwick (2018); CELLDAR/BAE.

Forward scatter: Abdullah et al. (2020); Musa et al. (2019); Babinet's principle.

Interferometry: Jian & Lu (2018); Doan et al. (2023); PMC11086345 (2024).

Acoustic: Noise2Weight (ResearchGate); PMC11946234 (2025); NEC Labs DAS (2022).

WiFi CSI: Di Seglio et al. (2024); Demissie (2025); UCL Chetty et al.