A hybrid artificial-intelligence system designed to authenticate radar-imaging signals from satellites recorded 96.9% accuracy on held-out test data, according to a new preprint. Its held-out macro-F1 score, a measure that balances performance across the two classes, was also 96.9%. QUASAR combines a conventional convolutional neural network with a variational quantum circuit, but that circuit was simulated on a classical computer rather than run on quantum hardware.
The result came from a 10% portion of a larger collection: 3.76 terabytes of raw IQ signal recordings gathered over 28 days from 37 operational ICEYE satellites using two USRP X310 receivers. The experiment therefore reports a result from one constellation and collection setup, not evidence that the system works across satellite networks generally.
What the test measured
The researchers converted the captured X-band synthetic-aperture-radar pulses into spectrograms, visual representations of how signal energy changes over time and frequency. Each one-versus-rest classifier used 2,021 balanced spectrograms, with whole orbital passes split into 60% for training, 20% for validation and 20% for testing.
QUASAR sends each spectrogram through a shared CNN encoder, then combines its classical representation with the output of an eight-qubit, four-layer variational circuit. The circuit uses amplitude-and-phase information from the complex IQ signal, and the two branches are fused before a binary classification decision.
The reported scores were averaged over 25 random initializations. The comparison models included a CNN-only system, a quantum-neural-network-only system, ResNet, MobileNet and a Transformer; the benchmarks used the same 10% subset, while the quantum circuit was evaluated through classical simulation.
Strong average scores, uneven satellite results
On the held-out test set, IQ-native QUASAR recorded 0.969 for both accuracy and macro-F1. The CNN-only comparator scored 0.894 for accuracy and 0.891 for macro-F1, an observed accuracy gap of 7.5 percentage points. The quantum-neural-network-only model scored 0.713 for accuracy and 0.709 for macro-F1.
The amount of data mattered most below the 10% mark: test accuracy was 71.1% with 5% of the corpus and 96.9% with 10%. Moving to 15% and 20% brought reported changes of only 0.6 and 0.9 percentage points from the 10% setting, respectively, within the study’s seed-to-seed variation.
The overall average concealed a wide spread between satellites. Of the 37 satellite enrollments, 28 had a reported hit rate of 1.00, while the other nine ranged from 0.93 down to 0.25; satellite X14 was the weakest in the reported results.
In a controlled encoding comparison, the IQ-native version scored 0.969 on test accuracy and macro-F1, against 0.947 for angle embedding. Its target recall was 0.975 versus 0.943, and it reached convergence at 37 rather than 68 epochs. The comparison does not establish that the encoding itself caused the difference.
Spoofing tests offer a mixed picture
The preprint reports rejection rates of 89.7% for replay, 94.1% for crafted-IQ injection and 81.3% for space-borne spoofing. The detailed evaluation supplied for the latter two scenarios does not include corresponding result tables or event counts, and the model variant associated with the summary is not fully reconciled with that caveat.
The most detailed attack result concerns laboratory replay signals from satellites X4, X20 and X41. Mean accuracy was 91.39%, macro-F1 was 0.9114, and recall was reported as 1.00 for all three satellites, while precision ranged from 0.769 to 0.952. The number of replay bursts was not reported.
What remains untested
The study is limited to one ICEYE constellation, one collection testbed and passively captured X-band imaging pulses. Its space-borne spoofing scenario used held-out ICEYE satellites as an emulation rather than measurements of a cross-constellation adversary, and the replay evaluation covered three satellites under laboratory conditions.
The evaluation used a representative 10% subset of the 3.76-terabyte corpus. The paper reports seed-to-seed variation but no confidence intervals, p-values, formal power calculation or attack-event counts, leaving the precision of the reported comparisons and spoofing rates less fully described.
Further testing would be needed on physical quantum hardware, other satellite constellations and hardware generations, and over long-term hardware drift. The reported results also do not show uniformly reliable authentication for every satellite: nine had degraded hit rates.
Exploratory analyses offered clues about what the classifier used. Gradient saliency on 50 correctly classified bursts concentrated on a contiguous window shorter than 1 millisecond at radar-pulse onset, while internal clustering measures were more favorable for QUASAR than for the CNN-only model. These analyses do not establish the physical mechanism behind the classifications.
A result still at the preprint stage
The manuscript was posted as arXiv:2608.20240v1 on 20 August 2026. Its appendix says the repository and dataset will be transferred to Zenodo upon acceptance.
Paper data and sources
Original title: QUASAR: A Quantum-Classical Neural Network for SAR Satellite Physical-Layer Authentication
Authors: Vincenzo Sammartino, Nathanael Denis, Roberto Di Pietro
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text