Keeping the Spectra Secret: Tensor-Split Physical-Layer Security for Hyperspectral Downlinks

A hyperspectral image records hundreds of narrow color bands per pixel. That makes it valuable, because the per-pixel spectrum identifies materials such as crop types, minerals, or camouflage paint, and it also makes it sensitive. When a satellite or drone sends such data over an open radio link, anyone with an antenna in range can listen. Physical-layer security is the family of methods that shape the transmitted radio signal with multiple antennas so that the intended receiver can decode a message while a listener at another location provably cannot. It is attractive because it needs no keys. Its weakness is throughput: the rate of bits that can be delivered secretly is often small, especially when the eavesdropper is close. A full hyperspectral cube is far too large to push through that narrow secret pipe.

This note proposes to stop treating the image as an undifferentiated block of secret bits. A widely used model for these cubes writes the scene as a few materials, each with one spectrum, laid out by a spatial map of where that material appears. The spectra are tiny, a few thousand numbers per scene, and they carry almost all of the material identity. The spatial maps are large but, on their own, say little about what anything is made of. We propose to send only the spectral factors through the secret channel and the spatial factors through an ordinary, high-rate public channel. The transmit beams are then designed by solving the joint non-convex problem of meeting a secrecy requirement on the small stream while delivering the large stream quickly. The central bet is that this split cuts the number of bits that must be secret by two to three orders of magnitude. If it does, physical-layer security becomes feasible for hyperspectral downlinks where it otherwise is not.

Core Claim

The proposal rests on the block-term model of Hyperspectral Super-resolution: A Coupled Nonnegative Block-term Tensor Decomposition Approach (CNN-BTD). That model writes the high-resolution cube $\mathcal{S}$ (width $I$ by height $J$ by $K$ bands) as

$$\mathcal{S} \approx \sum_{r=1}^{R} (\mathbf{A}_r \mathbf{B}_r^{T}) \circ \mathbf{c}_r ,$$

where $R$ is the number of materials, $\mathbf{A}_r\mathbf{B}_r^T$ is the rank-$L$ spatial abundance map of material $r$, $\mathbf{c}_r \in \mathbb{R}^K$ is its spectral signature, and $\circ$ is the outer product. The spectral payload is $KR$ numbers, roughly $4\times10^3$ for an Indian Pines-sized scene. The spatial payload is $(I+J)LR$ numbers, and the raw cube is $IJK$, about $4\times10^6$.

Our claim is that a confidential/public split along this factorization has two properties. First, the legitimate receiver can recover the cube with the same reconstruction quality as a full fusion pipeline. Second, an eavesdropper who decodes the public stream perfectly gains little material-level information, which we measure empirically. Third, total delivery time under a secrecy constraint falls sharply relative to securing the whole cube. Nothing in the research file couples the payload's tensor structure to secrecy beamforming. The closest work, Collaborative Compressive Detection With Physical Layer Secrecy Constraints, protects generic compressed measurements for binary detection using fixed artificial noise. It does not deal with image tensors, reconstruction, or optimized beams.

Research Program

The work has three parts.

On-board factorization. The platform runs coupled factorization on board. The fusion literature supplies this component: CNN-BTD for its physically interpretable spectra, and Hyperspectral Super-Resolution: A Coupled Tensor Factorization Approach (STEREO) for its identifiability guarantees, which show that the factors are unique even when the blur is unknown. The output is quantized factors $\{\mathbf{A}_r,\mathbf{B}_r\}$ and $\{\mathbf{c}_r\}$, plus an optional residual. The residual is a leakage risk, because it contains spectral information. We therefore either drop it or route it through the confidential stream, and we report both options.

Transmit design. The transmitter superimposes a confidential signal with covariance $\mathbf{Q}_c$ and a public signal with covariance $\mathbf{Q}_0$. We minimize the time $T$ to deliver one scene:

$$\min_{\mathbf{Q}_c,\mathbf{Q}_0 \succeq 0,\,R,\,b}\; T \quad \text{s.t.}\quad n_s(R,b) \le T\,R_s(\mathbf{Q}_c,\mathbf{Q}_0),\;\; n_0(R,b) \le T\,R_0(\mathbf{Q}_c,\mathbf{Q}_0),\;\; \mathrm{Tr}(\mathbf{Q}_c+\mathbf{Q}_0)\le P,\;\; d(R,b)\le d_{\max}.$$

In this problem, $n_s$ and $n_0$ are the secret and public bit counts as functions of the rank $R$ and the quantization bits $b$. $R_s$ is the secrecy rate of the confidential stream: the legitimate receiver's mutual information minus the eavesdropper's, each computed while treating the other stream as interference. $R_0$ is the rate of the public stream. $P$ is the power budget, and $d$ is the reconstruction distortion measured by spectral angle.

For fixed $(R,b)$, $R_s$ is a difference of concave functions, the standard secrecy structure. We handle it with successive convex approximation, which repeatedly linearizes the eavesdropper term. Where the confidential stream is single-beam, we use the semidefinite relaxation and the S-procedure robust formulation of Optimal and Robust Transmit Designs for MISO Channel Secrecy by Semidefinite Programming. That paper maximizes the secrecy rate of a single confidential message. Our problem adds a public stream and an outer search over $(R,b)$, so the payload's structure enters the constraints. The outer search is small and discrete, so we solve it by enumeration.

Eavesdropper leakage. We do not claim an information-theoretic guarantee about material inference from the public stream. Spatial shapes alone can hint at land use. Instead we attack our own scheme with strong eavesdroppers. One eavesdropper is a classifier trained on other scenes that predicts material classes from the spatial factors alone. A second matches the spatial factors against public spectral libraries to reconstruct a plausible cube.

Baselines

The naive reference is to send the full fused cube through a single confidential stream designed with the robust semidefinite method of Ma et al. This shows what securing everything costs in delivery time.

The strongest competitor is a "fuse on the ground" split. The low-resolution hyperspectral image goes through the secret channel, the high-resolution multispectral image goes in public, and the ground station fuses them with CNN-BTD. This is the natural design a fusion practitioner would choose, and it needs no on-board factorization. Its secret payload, however, is the whole low-resolution cube, and its public multispectral bands leak coarse spectra.

For legitimate reconstruction quality, the reference number is CNN-BTD's reported R-SNR of 27.17 dB and spectral angle of 0.0352 on Indian Pines (CNN-BTD), together with STEREO's 38.62 dB R-SNR on SALINAS (STEREO). Beating the baseline means matching these figures within 0.5 dB while cutting delivery time by at least an order of magnitude at equal eavesdropper leakage. For secrecy rate, Ma et al. report that the semidefinite design beats maximal-ratio transmission, but the file gives no table values. We therefore regenerate their synthetic Rayleigh setup ourselves rather than quote a figure.

All methods run on the same scenes, channels, power budget, and quantizer. Every method's beamformer is optimized with the same solver tolerance.

There are two ablations. The first isolates the beam design by using closed-form maximal-ratio or zero-forcing beams within our split. The second isolates the choice of split. It sends a random subset of factor entries of the same size as the secret payload, for example columns of $\mathbf{A}$ instead of the spectra, through the secret channel. If leakage under this random split is no worse than under the spectral split, the claim that spectra are the right thing to hide fails.

Why not just…?

Why not just encrypt with AES and skip physical-layer security? This is a real threat. Once our split shrinks the confidential payload to a few kilobits, encrypting it with AES is trivial. We regard the tensor split as the main contribution and treat physical-layer delivery as one way of protecting the small stream, not the only one. Physical-layer security still adds keyless protection that does not depend on the eavesdropper's computing power, which matters for platforms with fragile key distribution (Collaborative Compressive Detection). We report the split's benefit under both protection methods, so the result stands even if a reader prefers cryptography.

Why not just use maximal-ratio or zero-forcing beams? Ma et al. found that maximal-ratio designs lose secrecy rate and can fall to negative secrecy rates when the channel estimate is imperfect. Our first ablation measures how much of the delivery-time gain survives with these simple beams. If most of it does, we will say so plainly.

Why not just send the low-resolution hyperspectral image secretly and fuse on the ground? This is the strongest baseline above. It differs from our scheme in exactly two quantities: the size of the secret payload and the leakage from the public multispectral bands. The experiments measure both directly.

Proposed Experiments

Data and channels. The scenes are Indian Pines and SALINAS (airborne spectrometer data with roughly 200 bands) and CAVE (31 bands). We degrade each in the standard way to produce the low- and high-resolution inputs. Channels follow Ma et al.'s synthetic Rayleigh MISO wiretap model with 4 to 32 transmit antennas, sweeping transmit power and the channel-uncertainty radius.

Metrics. For the legitimate receiver we report R-SNR, spectral angle, and PSNR. For the system we report delivery time per scene and secret bits per scene. For the eavesdropper we report classification accuracy, using the ground-truth labels of Indian Pines and SALINAS, and the R-SNR of the library-matching attack.

Decisive checks. The proposal fails if any of three things happens. First, the eavesdropper's classification accuracy from the public spatial factors comes within a few points of its accuracy from the public multispectral image in the ground-fusion baseline. Second, the random-split ablation leaks as little as the spectral split. Third, the legitimate receiver's spectral angle degrades beyond CNN-BTD's 0.0352 at the rank and quantization the optimizer selects.

We also test robustness under imperfect eavesdropper channel knowledge using the S-procedure formulation. The non-robust variant is expected to show negative secrecy rates, which is the failure mode Ma et al. describe. We will check whether it does.