X-band SAR satellites (8-12 GHz) play a critical role in disaster response, environmental monitoring, and military intelligence. Yet, they lack robust physical-layer authentication (PLA), a security layer orthogonal to cryptographic solutions. Existing PLA systems, typically based on radio-frequency fingerprinting, are often limited to sub-6 GHz frequencies and rely on classical deep learning. However, this approach underfits the IQ phase nonlinearities that distinguish satellite hardware. In this paper, we present QUASAR, to the best of our knowledge the first quantum-classical hybrid architecture that fuses a CNN spectrogram encoder with a variational quantum circuit (VQC) to provide PLA to X-band SAR signals. Our solution enjoys two distinctive features: (i) it is markedly more data-efficient than classical machine learning, requiring only 10% of the training data to match the accuracy of classical baselines -- data collection being notoriously the most time-consuming phase of PLA; and, (ii) at an equal data budget, it improves classification accuracy over those baselines. In detail, we test our solution under three adversarial scenarios: replay, crafted-IQ injection, and space-borne spoofing. QUASAR rejects spoofed transmissions in 89.7%, 94.1%, and 81.3% of attempts, respectively, establishing the first quantum-enhanced physical-layer classifier for satellite constellations. The fully detailed framework and the supporting results, other than being interesting on their own, show a novel research avenue for physical-layer authentication.
Figures & tables
Fig. 1: IQ samples and qubit states. Left: a complex IQ sample x=I+jQ , represented as a vector in the complex plane with magnitude r and phase φ . Right: a single-qubit pure state ∣ψ⟩=cos(θ/2)∣0⟩+eiφsin(θ/2)∣1⟩ on the Bloch sphere, parameterized by the polar angle θ and the azimuthal angle φ . Both objects are fully specified by two real parameters: the IQ magnitude r maps to the polar angle θ=2arcsin(r) , while the phase φ is the same physical quantity in both representations.
Fig. 2: Threat model. (A) A legitimate SAR satellite Xt transmits two physically distinct X-band signals: i) (frequent) high-power chirp pulses for Earth imaging; and, ii) (uncommon) phase-modulated data downlinks. (B) A ground station captures imaging pulses for authentication purposes. Two adversaries are considered: (C) a space-borne adversary operating a SAR satellite Xs of comparable hardware; and, (D) a ground spoofer operating a commercial high-end SDR.
Fig. 3: Setup to collect satellite signals in the X-band. Two USRPs are used, along with an MX12000 downconverter and 8TB Seagate BarraCuda storage.
Fig. 4: Time-domain IQ waveforms (left) and constellation density maps (right) for three ICEYE satellites: X11 (top), X13 (middle), and X14 (bottom).
Fig. 5: QuaSAR end-to-end architecture. The shared CNN encoder (blue) maps the 224×224 STFT spectrogram to a 256-dimensional latent vector h . The QNN branch (violet) projects h to 8 complex components, loads each onto one qubit through the IQ-native amplitude–phase encoding of Eq. ( 6 ), applies L=4 strongly entangling layers (SEL), and emits 8 Pauli-Z expectation values. The classical skip (green) compresses h through two FC layers to 64 dimensions. Late fusion concatenates both into a 72-dimensional vector classified by a linear head.
Fig. 6: Test accuracy of QuaSAR as a function of the training-data fraction. The fraction is the share of the full balanced corpus drawn by stratified random sampling, before the train/validation/test split is applied. For each fraction, the 60/20/20 split is rebuilt, and all hyperparameters are kept fixed. Each marker reports the mean over 25 random seeds.
Angle embedding
IQ-native (ours)
Metric
Validation
Test
Validation
Test
Accuracy
0.950 ( ± 0.010)
0.947 ( ± 0.009)
0.973 ( ± 0.010)
0.969 ( ± 0.009)
Macro-F1
0.950 ( ± 0.004)
0.947 ( ± 0.005)
0.973 ( ± 0.004)
0.969 ( ± 0.005)
Precision (target)
0.953
0.951
0.969
0.965
Recall (target)
0.948
0.943
0.974
0.975
Convergence epoch
68
—
37
—
TABLE I: Binary authentication performance under the two quantum embeddings, all other components of the architecture held fixed. Angle embedding (Eq. ( 3 )) loads a single real projection per qubit; the IQ-native encoding (Eq. ( 6 )) loads amplitude and phase. Mean ± std. dev. over 25 random seeds.
Fig. 7: Per-satellite hit rate (circles) and miss rate (crosses) for the full 37-satellite ICEYE constellation, sorted by descending hit rate. Twenty-eight satellites achieve hit rate =1.00 and miss rate =0.00 ; a group of nine satellites exhibits degraded performance, with X14 reaching the lowest hit rate (0.25).
Model
Accuracy
Macro-F1
Ablation variants
QNN-Only (PCA + VQC)
0.713
0.709
CNN-Only (no VQC)
0.894
0.891
QuaSAR (angle embedding)
0.947
0.947
Classical spectrogram baselines
Spectrogram ResNet
0.854
0.861
TABLE II: Ablation and benchmark results on the binary authentication task ( Dtest , mean over 25 seeds). The upper block reports ablation variants of QuaSAR ; the lower block reports external classical spectrogram classifiers trained under the identical protocol.
Satellite
Acc
Prec
Rec
F1
X4
0.85
0.76
1.00
0.86
X20
0.91
0.80
1.00
0.88
X41
0.97
0.95
1.00
0.97
Mean
0.91
0.84
1.00
0.91
TABLE III: Per-satellite spoofing detection performance under Scenario (A) replay. Spoofed IQ signals from three operational ICEYE satellites are evaluated against the trained QuaSAR classifier. Recall = 1.00 on all satellites confirms zero missed detections.
Fig. 8: t-SNE projection of the 72-dimensional fused latent space z for three ICEYE satellites (X4, X20, X41). Filled circles denote legitimate bursts. Crosses denote spoofed bursts generated under Scenario (A) replay. Each satellite identity occupies a distinct, compact region; spoofed samples are displaced from their legitimate counterparts in all three cases, confirming that the quantum-classical encoder preserves and amplifies inter-class fingerprint distance.
Model
Silhouette ↑
Calinski–Harabasz ↑
Davies–Bouldin ↓
CNN-Only
0.3106
61.69
1.2219
QuaSAR
0.3775
64.18
1.0341
Δ (rel.)
+21.6%
+4.0%
−15.4%
TABLE IV: Internal clustering quality of the 72-dimensional fused latent space z for QuaSAR and the CNN-Only baseline. Higher Silhouette Score and Calinski–Harabasz index indicate better-separated clusters; lower Davies–Bouldin index indicates more compact clusters. All indices are computed on the full test partition without dimensionality reduction.
Fig. 9: Decision surface for satellite X41 in the PCA-projected latent space z . The color map encodes the posterior P(spoof∣z) fitted by a logistic classifier on the two leading principal components; the dashed line marks the P=0.5 iso-contour. Legitimate bursts (blue circles) cluster in the low-probability region; spoofed bursts (red crosses) concentrate in the high-probability region, with no cross-boundary misclassification for X41.
TABLE V: Positioning of QuaSAR against representative prior work in satellite RF fingerprinting, physical-layer authentication, and quantum-hybrid RFFI. Rows are grouped by methodological family.