Physical Trust from Mutual mmWave Radar Interference

Most radar systems are designed to fight interference. This research asks what happens if we stop treating all interference as waste. At millimeter-wave frequencies, an interfering FMCW signal does not travel through empty space: it propagates through the same walls, objects, reflectors and geometry that define the radio environment around the two radars. That propagation leaves structure in the received complex signal. The project investigates whether that structure can become a source of shared physical trust.

The concept is simple to state and difficult to engineer. Two independent radar nodes observe the same bistatic environment from opposite ends of a link. They do not share an RF clock, they do not exchange dedicated probing pilots, and in the experimental zero-message mode they do not reconcile their bits afterwards. Each radar must decide locally what it trusts, which multipath components are physically meaningful, which phase relationships are stable enough to quantify, and when it should refuse to generate a bit at all.

The objective is not to make random numbers from interference. It is to isolate the part of the interference that belongs to the shared physical channel.

That distinction is the core of the project. A security system built on raw phase would fail almost immediately because raw phase is contaminated by local oscillators, PLL behavior, timing offsets, RF-chain phase, antenna geometry and noise. The interesting question is whether those non-reciprocal terms can be reduced enough to expose the underlying propagation relationship that both legitimate nodes observe.

01Experimental basis

The experimental platform is built around Texas Instruments IWR1443BOOST mmWave radars operating with an effective sweep in the 77–81 GHz region. One transmitter and four physical receivers are active, giving a SIMO measurement geometry and enough spatial diversity to form multiple differential observations across the receive array.

The FMCW waveform itself is conventional. An ideal chirp can be written as

with instantaneous frequency \(f_{\mathrm{TX}}(t)=f_0+St\). What matters for this research is what happens when one radar receives another radar's chirp through a physical propagation path. In a bistatic geometry, the measured delay corresponds to the total transmitter-to-scatterer-to-receiver path length \(D\):

Because the radars are independent, their chirps do not line up permanently. They drift in and out of overlap as their timing evolves. That is not a defect in the concept. It is precisely why mutual interference appears intermittently and why source identification, causal tracking and temporal validation matter. In the pseudo-parallel interference regime, the interfering component can move through the delay profile in a structured way, allowing the receiver to isolate physically meaningful bistatic paths instead of treating the event as undifferentiated noise.

After range-domain processing, this research works with a complex measurement \(X_s[t,k,r]\), where \(s\) is the observing radar, \(t\) is time, \(k\) is a delay bin and \(r\) identifies one of the four receivers. For a tracked path \(p\), the corresponding complex sample is

The phase \(\phi_{s,p,r}(t)=\arg Z_{s,p,r}(t)\) is useful, but not directly trustworthy. A practical measurement can be modeled as

where \(H^{\mathrm{phys}}\) is the physical channel, \(G_s\) represents the local complex hardware response, \(\eta_s\) groups timing and oscillator-related phase terms, and \(W_s\) is noise. The physical channel may be reciprocal while the complete measurement is not research therefore does not search for equality of raw complex samples; it searches for reciprocity in carefully constructed differential observables.

02Differential phase

The project treats phase as what it really is: a circular variable. Subtracting \(179^\circ\) and \(-179^\circ\) as ordinary numbers is meaningless for a radio phase estimator. Every physical phase difference is wrapped onto the circle,

and stability is measured with circular statistics. For a set of phase samples,

The concentration \(\rho\) tells us whether a candidate phase relationship is locally stable. A feature with \(\rho\approx1\) is tightly clustered; a feature with low \(\rho\) may be noisy, multimodal or physically mixed. An associated circular dispersion is

These quantities are not cosmetic diagnostics. They decide whether a physical observation is stable enough to be allowed near a quantizer.

The most important step is differential processing. If \(\tilde{\phi}_s(t,p,r)\) denotes a causally cleaned phase, this research can compare two receivers observing the same path,

or compare a multipath component against the LOS/reference path on the same receiver,

The strongest physically interpretable combination is the double difference,

In complex form, the same quantity can be obtained directly as

This is important both physically and computationally. It avoids unnecessary unwrap/re-wrap operations and allows much of the live extractor to remain in the complex domain until the final decision stage.

Four receivers give six nominal RX pairs, but only three independent linear spatial differences:

That may sound like a small detail, but it is exactly the kind of detail that separates a physically defensible key-generation experiment from an entropy overestimate research does not assume that every available measurement dimension is independent just because the software can compute it.

03Multipath and selection

Multipath is where the project becomes particularly interesting. The environment creates multiple reflected paths, each with its own delay and phase. Those paths are a source of spatial diversity, but they are also a major source of failure if two contributions fall inside the same delay bin.

If a measured bin contains

its observed phase can move simply because the relative amplitudes \(A_1/A_2\) change, even if the underlying phases remain constant. That can create apparently rich phase dynamics with no useful reciprocal meaning.

For this reason this research does not select bits only from “strong peaks”. It measures whether the local energy is dominant, whether nearby peaks are sufficiently separated, whether the track is stable, whether the phase remains concentrated, and whether the candidate stays far enough away from the quantizer boundary.

The system is deliberately allowed to abstain. That matters because in a zero-message configuration there is no reconciliation step to repair a wrong bit. A questionable cell is more dangerous than a missing cell.

Once a feature survives the physical gates, the binary decision can be written as

A guard band removes observations close to the threshold. The purpose is not to maximize the number of emitted bits; it is to maximize the probability that two legitimate nodes independently make the same decision from their own local measurement.

The design philosophy is simple: never interpret a bit before explaining the physical path that produced it.

04Local trust

This is where this research becomes relevant to security rather than only signal processing. The system does not compare Alice and Bob first and then search for a feature that happens to match. Each node must produce its candidate independently:

Only afterwards are candidate blocks compared. A third observer is evaluated afterwards as well. That separation prevents the extractor from “learning the answer” from Bob or Eve.

The project has demonstrated that this physical approach is viable: independently observed bistatic interference can contain enough reciprocal structure for legitimate radar pairs to derive matching short candidate material under appropriate channel conditions and quality control. That is the essential capability this research is built around.

The research now focuses on making that capability more robust, more observable in real time, and more rigorously tied to physical path identity. The real-time architecture uses directed source-separated channels, causal path tracking, four-RX complex processing, live quality metrics and digest-based comparison so that the extraction process remains local to each radar.

With three nodes \(R_0,R_1,R_2\), the system does not treat “the interference” as one undifferentiated signal. It separates directed relationships such as

because a trust relationship between \(R_0\) and \(R_1\) is meaningful only if the underlying observation actually belongs to that pair. Source separation is therefore part of the security model, not just part of the radar visualization.

05Security applications

The broader idea reaches beyond radar. A conventional credential tells a system who possesses a key. A physical-layer observation can potentially add a second question: are these two devices currently embedded in the physical relationship we expect them to share?

That is why this research is interesting for defense robotics, cooperative sensing, tactical systems and high-assurance infrastructure. The project is not proposing to replace AES, PKI, secure elements or authenticated protocols. It is exploring a new signal that those systems could potentially consume.

For autonomous robotic systems, that could mean adding a physical factor when two nodes establish local trust. For distributed radar, it could mean verifying that cooperating sensors continue to observe a compatible bistatic relationship. For critical infrastructure, it could provide another environmental signal before granting access to protected processing or data. For military communications, it could eventually contribute additional physical entropy or channel-binding evidence alongside conventional cryptographic protocols.

Those are future uses, not claims of current operational readiness. This remains a research platform, and that is precisely why the next stage matters.

06Next research stage

The project is now moving from proof of concept toward a more complete experimental security platform. The priority is not to chase larger bit counts. The priority is to understand exactly when the physical channel should be trusted and when the extractor should remain silent.

That requires better live path identity, stronger multi-radar source separation, richer diagnostics, strict capture traceability, freshness and temporal-entropy analysis, more controlled spatial measurements, and eventually active-adversary testing. It also requires bridging several disciplines that are rarely developed in the same stack: FMCW radar processing, mmWave channel modeling, physical-layer security, statistical inference and cryptographic entropy analysis.

This is also where external collaboration can have disproportionate value. Additional radar hardware, controlled RF chamber access, wavelength-scale positioning systems, synchronized instrumentation for validation, specialist cryptographic review, adversarial RF expertise and autonomous-platform testbeds would all accelerate the research.

There is a broader reason to pursue this line of work. As sensing platforms become more distributed and more autonomous, security cannot rely only on what devices claim to be. There is value in asking what the physical environment can tell us about the relationship between them.

This work attempts to turn that idea into an engineering discipline: from mutual interference, to bistatic channel measurement, to differential phase, to causal physical validation, and finally to key material that exists because two nodes really did observe the same radio world.

The ambition is not to make interference disappear. It is to make part of it useful.

This is an ongoing experimental research project. It does not currently claim certified cryptographic security, formal information-theoretic secrecy, active-adversary resistance or deployment-ready military key management. Its present contribution is a physically grounded and experimentally validated path from mutual mmWave FMCW interference to reciprocal candidate key material, with a real-time implementation under active development.