arXiv:2610.13211
Verifiable Sampling for Delegated Generative Quantum Machine Learning
Matteo Inajetovic, Anna Pappa, Petros Wallden
Abstract
Generative machine learning has shown remarkable results across a wide range of interdisciplinary tasks, motivating significant efforts toward Quantum Generative Models as a path to beyond-classical performance, given their capability of generating classically intractable distributions. However, comparatively little attention has been paid to securely delegating such computations to a powerful remote quantum server. In this work, we propose a blind and verifiable protocol for sampling that can be used for the inference phase of such quantum models. Such a protocol inherits blindness from Universal Blind Quantum Computing and achieves verifiability by exploiting known approaches based on trapification schemes. Alongside, to delegate the training phase, we discuss an extension of the previously proposed verifiable end-to-end Delegated Variational Quantum Algorithms (Inajetovic et al., Phys. Rev. Research, 2026). Beyond the usual Accept/Abort outcome characterizing verification protocols, the sampling protocol for inference also provides, even against a malicious server, an estimation of the amount of corrupted generated samples, i.e., a fine-grained feature describing the quality of the outcome connecting verification protocols and application-level benchmarking.
AI Summary
The paper studies blind, verifiable delegation of sampling from quantum generative models, together with extensions to delegated training. It combines UBQC blindness with existing trapification schemes: computation and test rounds are randomly interleaved, and sampling is accepted only if no trap fails. The client must prepare and send quantum states; security is intended to be information-theoretic against a malicious server. The analysis connects a corrupted computation-round fraction $\gamma$ to distributional error through the mixture $\hat p=(1-\gamma)p+\gamma q$. It derives $\mathrm{MMD}_k(p,\hat p)\leq2\sqrt K\gamma$ for kernels satisfying $k(x,x)\leq K$, total-variation error at most $\gamma$, and $D_{\mathrm{KL}}(p\|\hat p)\leq-\log(1-\gamma)$. Using prior trap-verification arguments, it states exponentially decreasing probability of accepting an output exceeding a chosen error threshold. It also proposes estimating corruption from trap failures, with confidence-interval half-width $c\sqrt{\ln(2/\alpha)/(2t)}$, where $c$ is the graph-coloring number and $t$ the test-round count; this derivation models detections as independent binomial trials. Training is discussed through bounds relating corrupted samples to gradient error, illustrated with QCBMs and an MBQC generative model. Numerical demonstrations use a four-qubit, $2\times2$ Bars-and-Stripes task under single-output-qubit dephasing attacks, including sweeps with 500 computation rounds and 100 or 500 tests. The analysis neglects finite-shot distribution-estimation error, leaving its incorporation for a fuller treatment. Shared-trap training requires identical MBQC graphs across gradient evaluations; the illustrated relative-gradient bound additionally assumes a nonzero gradient-norm lower bound. Noise-tolerant acceptance thresholds are left for future work.
Central claim may not hold: The adversarial corruption-rate confidence interval assumes an exact trap-detection model that blindness and trapification do not establish, so its advertised coverage can fail. The sampling theorem does not supply a finite-shot empirical-distribution guarantee beyond its explicitly assumed negligible-shot-noise regime. The training example incorrectly applies a parameter-shift identity directly to the nonlinear, unsquared MMD loss. The training example incorrectly converts per-evaluation corruption bounds into a bound using the global corruption fraction.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 6 pages, 1 figure, 1 table. Accepted at the 2026 IEEE 2nd International Conference on Quantum Artificial Intelligence (QAI 2026). \c{opyright} 2026 IEEE