acceptodds
Under review as a conference paper at ICLR 2027

Privacy Aware Quantum–Classical In-Context Learning for Tabular Data

Abstract

Remote quantum serving for tabular in-context learning must simultaneously address privacy-sensitive context data, variable-sized task inputs, communication overhead, and tight quantum-resource constraints. We propose QICL, a privacy-aware split quantum-classical framework built around a compact structured interface between the client and a remote quantum server. On the client side, QICL maps row-level contextual representations into bounded class-wise summaries and combines them with class-prior information to form a compact interface with explicitly controlled record-level sensitivity. A block-wise Gaussian mechanism privatizes this interface before transmission, while the interface changes the dominant context communication from record-wise to class-wise scaling. On the server side, a candidate-aware relational predictor constructs query-candidate-rival interactions before compressing them into quantum-compatible coordinates for few-qubit prediction. We establish record-level differential privacy for the uploaded privatized interface and derive a condition linking privacy perturbation to prediction stability. Experiments on TALENT classification datasets evaluate QICL against multiple tabular ICL models, strong conventional tabular predictors, and classical prediction heads under the same structured interface, demonstrating a favorable privacy-utility-communication trade-off under constrained quantum resources.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Quantum Machine Learning, Tabular In-Context Learning, Differential Privacy, Hybrid Quantum-Classical Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8cNLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies tabular in-context learning in a setting where the client keeps its data local and hands the prediction step to a remote quantum server. In this work, they propose QICL, which is a privacy-aware split quantum–classical framework. This framework compresses labeled context records into bounded class-wise summaries and class proportions, applies a classical Gaussian mechanism to provide record-level differential privacy (DP), and combines classical query-candidate-rival comparisons with a quantum predictor.

Critical strengths
  • The problem framing is clear, addressing privacy, communication costs, and limited quantum resources.
  • A formal DP guarantee is included to establish record-level privacy protection for the uploaded context information.
  • The experiments are broad, including 200 datasets, multiple baselines, component ablations, and real-hardware execution.
Critical weaknesses
  • My first concern is about the necessity of the quantum component and the limited novelty of the analysis. Specifically, the manuscript does not convincingly establish what this quantum component contributes beyond a suitably trained classical predictor. The theoretical contribution is also limited. The privacy guarantee largely applies bounded-sensitivity analysis and the classical Gaussian mechanism, while prediction stability follows from Gaussian concentration and a Lipschitz margin argument. Neither result uses any property of the quantum circuit. They support the design, but they add little methodological novelty and do not explain why quantum computation is needed.

  • My second concern is about the practical choice of nmin⁡n_{\min}. The theory requires this threshold to be fixed independently of private data. However, overestimating the minimum class size can produce insufficient noise for the stated privacy guarantee, while underestimating it can introduce excessive noise and hurt prediction accuracy. The paper does not report the values used or explain how they are selected under the independence requirement. This makes it difficult to verify whether the experiments satisfy the theoretical assumptions and whether the reported privacy–utility trade-off is justified.

  • My third concern is about whether the experiments show the usefulness of the quantum component. In Figure 6, the quantum predictor offers only a small improvement over classical heads, even in the simulation setting. The hardware evaluation in Appendix even shows a substantial performance drop. Accuracy falls from 75.1% in simulation to 50.0% on hardware for the same nine datasets. These results again leave unclear why a quantum component is needed and what practical benefit it provides.

Questions for authors

See the weaknesses above.

9 comments
AuthorsOpenReview ↗

Response to Reviewer 8cNL (1/3)

We thank the reviewer for the careful assessment and for recognizing the breadth of our experiments. We address the three concerns below and all changes are marked in blue in the revised PDF.

Concern 1. Necessity of the quantum component and novelty of the analysis. We thank the reviewer. We first clarify the objective, then address the empirical contribution of the quantum component, the necessity of quantum computation in our setting, and the scope of the theory.

(1) Objective. QICL does not claim a general quantum advantage. It asks a narrower question: what client--server interface makes few-qubit serving of tabular ICL viable under joint privacy, communication, and qubit constraints?

(2) Empirical contribution of the quantum predictor. Under the identical privatized interface (Figure 6), paired Wilcoxon signed-rank tests across 200 TALENT datasets with Holm correction give adjusted pp-values of 4.13×10−54.13\times10^{-5}, 4.91×10−74.91\times10^{-7}, and 4.37×10−74.37\times10^{-7} against MLP, Fourier, and Transformer heads. All are significant. The gains are modest but consistent, so the improvement cannot be attributed to the interface alone.

(3) Why quantum computation is considered. Our motivation is deployment, not advantage. QPUs are accessed through remote cloud services [1], unlike classical models that run on local infrastructure. This creates a client--server boundary across which sensitive task information must pass, making record-level representations undesirable due to privacy leakage, communication overhead, and the limited input capacity of few-qubit circuits. Few-qubit predictors are of interest because data re-uploading circuits realize trigonometric models whose frequency spectrum grows with encoding repetitions [2,3], yielding expressive nonlinear predictors with few parameters; rigorous quantum--classical separations exist for specific structured problems [4]. Whether these extend to tabular prediction is open, and we make no such claim.

(4) Scope of the theory. The upload-level DP guarantee deliberately relies on established Gaussian mechanisms and sensitivity analysis: because the client transmits classical representations to a remote QPU, privacy must be established before transmission and cannot depend on quantum computation that occurs after it. The analysis is nonetheless specific to the QICL interface: its bounded structured representation enables explicit record-level sensitivity calibration, and the stability analysis characterizes how privacy perturbations propagate through candidate-aware inference. The contribution is the integration and formal characterization of privacy within a communication-efficient quantum--classical architecture.

We further extend the theory to the quantum model. Under a binary Pauli-measurement model with depolarization probability pp and NshotN_{\mathrm{shot}} independent measurement shots, if the uploaded privatized interface satisfies (ϵ,δ)(\epsilon,\delta)-DP, the released binary-measurement outputs satisfy (ϵq,δq)(\epsilon_q,\delta_q)-DP, with \begin{equation*} \epsilon_q =\log\left( \frac{ e^\epsilon+ \left(\frac{p}{2-p}\right)^{N_{\mathrm{shot}}} }{ 1+e^\epsilon \left(\frac{p}{2-p}\right)^{N_{\mathrm{shot}}} } \right), \qquad \delta_q

\frac{ 1-\left(\frac{p}{2-p}\right)^{2N_{\mathrm{shot}}} }{ 1+e^\epsilon \left(\frac{p}{2-p}\right)^{N_{\mathrm{shot}}} }\delta. \end{equation*} This bounds the contraction of privacy loss induced by the quantum model. It does not replace upload-level DP: the honest-but-curious server sees the uploaded representations and remains covered by the original guarantee. The tighter bound applies to an output-only adversary observing the specified measurement record or its post-processing. Appendix C.4 extends the bound from a single binary Pauli observable to the full predictor, which measures all nqn_q qubits over multiple candidate circuits. Since the additional amplification may be negligible at large shot counts or low depolarization probabilities, we present this result as a complementary characterization of output-level privacy rather than the primary privacy mechanism.

(5) Privacy–utility trade-off. To quantify the privacy–utility implications of quantum noise, we investigate the privacy--utility trade-off at a fixed output-level privacy budget of ϵq=1\epsilon_q=1. Lower depolarizing probabilities and more measurement shots improve prediction accuracy by reducing channel-induced distortion and measurement estimation variance, respectively. These results also suggest that calibrated client-side interface perturbation provides a more favorable privacy--utility trade-off than relying on quantum execution noise for privacy protection.

The revised manuscript incorporates the statistical tests (Table 4), the deployment motivation (Sections 1 and 2) and scope of the theoretical claims (Section 3.5 and Appendix C.4), and the quantum noise analysis with experiments (Figure 8 and Appendix D. 4).

AuthorsOpenReview ↗

Response to Reviewer 8cNL (2/3)

Concern 2. Practical choice of nmin⁡n_{\min}. We thank the reviewer, and we apologize that the submitted manuscript omitted these details. We presented the sensitivity analysis in its simplified minimum-support form, whereas all reported experiments use a threshold computed from privatized quantities that enforces the bound by construction. We now document the detailed procedure in Appendix D.1.

(1) Threshold selection. With πc=nc/Nctx\pi_c=n_c/N_{\text{ctx}}, the client first privatizes the class prior and computes

n~c=Nctxπ~c,n^min⁡=max⁡(n0, ⌊[min⁡c∈Yn~c]+⌋),\tilde n_c=N_{\text{ctx}}\tilde\pi_c,\qquad \hat n_{\min}=\max\Big(n_0,\ \big\lfloor[\min_{c\in\mathcal Y}\tilde n_c]_+\big\rfloor\Big),

where [x]+=max⁡(0,x)[x]_+=\max(0,x) and n0=2n_0=2 is a predefined public lower bound. The threshold is a post-processing of π~\tilde\pi and does not directly use the unperturbed class counts for threshold selection. NctxN_{\text{ctx}} is invariant under replace-one adjacency, and Y\mathcal Y is fixed by the task schema.

(2) Bound holds by construction. The summary uses a clamped denominator,

Sc(n^min⁡)=1max⁡(nc,n^min⁡)∑i:yi=cψper(hi).S_c^{(\hat n_{\min})}=\frac{1}{\max(n_c,\hat n_{\min})}\sum_{i:y_i=c}\psi_{\text{per}}(h_i).

Since ∥ψper(hi)∥2≤1\lVert\psi_{\text{per}}(h_i)\rVert_2\le1, we have ΔS(n^min⁡)≤22/n^min⁡\Delta_S(\hat n_{\min})\le2\sqrt2/\hat n_{\min} for every value of n^min⁡≥1\hat n_{\min}\ge1 and every neighboring pair, including when a class has fewer than n^min⁡\hat n_{\min} records. Because n^min⁡\hat n_{\min} depends on π~\tilde\pi, the client privatizes the prior block first and then, conditionally on π~\tilde\pi, privatizes the summary block with σS\sigma_S calibrated so that ΔS(n^min⁡)/σS\Delta_S(\hat n_{\min})/\sigma_S equals the constant in Theorem 1. By adaptive composition of Gaussian mechanisms, this sequential release satisfies the same (ϵ,δ)(\epsilon,\delta)-DP as Theorem 1, so selecting n^min⁡\hat n_{\min} from π~\tilde\pi incurs no additional privacy cost (Appendix D.1).

(3) Neither failure mode violates the guarantee. If n^min⁡\hat n_{\min} overestimates a class size, that class's summary is divided by n^min⁡\hat n_{\min} instead of ncn_c and shrunk toward the origin; the sensitivity bound still holds, and the scale-invariant cosine similarities used for rival selection are unaffected. If n^min⁡\hat n_{\min} underestimates the minimum class size, the discrepancy arises only from prior-block noise and is smaller. Our rule is a best-effort estimate from privatized data alone; public, non-sensitive class-size knowledge (e.g., population proportions or domain priors) can sharpen n^min⁡\hat n_{\min} and improve the privacy--utility trade-off without affecting the guarantee.

AuthorsOpenReview ↗

Response to Reviewer 8cNL (3/3)

Concern 3. Usefulness of the quantum component. We address the simulation results, the hardware results, and the practical benefit separately.

(1) Simulation. The quantum predictor's advantage is small but statistically consistent, and we do not interpret it as a general quantum advantage. Under the identical privatized interface, paired Wilcoxon signed-rank tests across the 200 TALENT datasets with Holm correction give adjusted pp-values of 4.13×10−54.13\times10^{-5}, 4.91×10−74.91\times10^{-7}, and 4.37×10−74.37\times10^{-7} against the MLP, Fourier, and Transformer heads (Table 4). A few-qubit predictor is therefore at least as effective as classical heads of comparable size, with a small but consistent improvement.

(2) Hardware. We agree that the original drop from 0.7510.751 to 0.5000.500 was substantial: the model had been trained in noiseless simulation and executed directly on hardware without adaptation to device noise. We have therefore performed noise-aware training. We obtained the device-calibrated noise model of ibm_kingston, ran additional local noise simulations of hardware imperfections with 64 shots, and re-evaluated QICL on the real ibm_kingston processor with the same nine datasets, 6464 shots, and the same privacy budget (ϵ=1\epsilon=1, δ=10−5\delta=10^{-5}), retaining the client-side Gaussian mechanism. Average hardware accuracy rose from 0.5000.500 to 0.7080.708, a 20.820.8 percentage-point improvement, with the residual gap attributable to finite-shot estimation, gate and readout errors, and mismatch between the noise model and the device at execution time. QICL thus retained performance close to its ideal-simulation counterpart under the same privacy setting. Quantum noise experiments (Figure 8 and Appendix D.4) further show that more shots and lower depolarizing probability improve accuracy, confirming how quantum noise affects prediction.

(3) Practical benefit. These results establish the feasibility of combining a compact differentially private interface with a few-qubit predictor executable on existing hardware: classical comparisons show a modest predictive benefit, and hardware experiments show this performance was largely retained on-device. We do not claim that QPU inference is more efficient end to end than classical inference. The value of QICL is that it addresses a setting in which a remote quantum backend is used and supplies what such deployment requires: a record-level DP guarantee for the uploaded context that holds regardless of server behavior, communication scaling with the number of classes rather than context records, and an input compact enough for few-qubit circuits. QICL is thus a privacy-preserving serving architecture usable now and applicable as hardware matures.

We have updated limitations and future work in Section 5 to state that a residual simulation-to-hardware gap remains after noise-aware training, to be addressed by hardware-informed calibration, error mitigation, and hardware-in-the-loop fine-tuning.

References

[1] Nguyen, H. T., Krishnan, P., Krishnaswamy, D., Usman, M., & Buyya, R. (2024). Quantum cloud computing: A review, open problems, and future directions. arXiv preprint arXiv:2404.11420.

[2] Pérez-Salinas, A., Cervera-Lierta, A., Gil-Fuster, E., & Latorre, J. I. (2020). Data re-uploading for a universal quantum classifier. Quantum, 4, 226.

[3] Schuld, M., Sweke, R., & Meyer, J. J. (2021). Effect of data encoding on the expressive power of variational quantum-machine-learning models. Physical Review A, 103(3), 032430.

[4] Liu, Y., Arunachalam, S., & Temme, K. (2021). A rigorous and robust quantum speed-up in supervised machine learning. Nature physics, 17(9), 1013-1017.

Reviewer_3kzhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QICL is a split quantum–classical framework for private tabular in-context learning. A frozen TabICL front-end processes records locally; bounded multi-harmonic class summaries and class priors are privatized with a jointly calibrated Gaussian mechanism before upload. The server forms query–candidate–rival features and scores candidates through a shared data-reuploading quantum circuit with a classical readout. The privacy results are conditional on record-level sensitivity assumptions; prediction stability additionally assumes a fixed rival ordering and Lipschitz control. Across 200 TALENT datasets, mean accuracy is 0.532 at ε = 1 and 0.548 at ε = 4, with communication scaling of 73C + 72Nq. The study includes five-seed results, paired tests, ablations, and a nine-dataset hardware experiment. (Sections 3.3–3.5; Tables 1–3; Figures 4–9; Appendices C–D.)

Critical strengths
  1. The interface links communication reduction to an explicit privacy mechanism. Unit-norm multi-harmonic features, bounded class aggregation, and block-wise Gaussian release support a record-level sensitivity analysis. The guarantee is scoped to context records rather than stated as a general privacy claim. (Equations 1–6; Proposition 1; Theorem 1; Table 2.)
  2. The 200-dataset evaluation is broader than a few favorable examples. It reports several privacy budgets, five seeds, paired Wilcoxon tests with Holm adjustment against the listed main baselines, and ablations of summaries, rivals, and qubit count. (Section 4; Table 3; Figures 4–9.)
Critical weaknesses
  1. The quantum predictor’s contribution is small under the same interface: approximately 0.532 accuracy versus 0.528 for the classical Transformer head. Table 3 does not test this head-level contrast. On the nine hardware datasets, accuracy falls from 0.751 in simulation to 0.500 with 64 shots. These results demonstrate executability, but leave practical utility uncertain without latency and execution-cost measurements. (Figure 6; Section 4.2; Appendix D.3.)
  2. The privacy comparison is difficult to audit. Theorem 1 requires a non-private minimum class-support threshold, but the concrete threshold and checks across TALENT tasks are not reported. Conventional predictors receive perturbed input data, while QICL releases a bounded aggregate. The baseline sensitivities and noise calibration need enough detail to establish that the nominally shared ε values support comparable guarantees. (Sections 3.5.1 and 4.1; Appendix C.)
Questions for authors

What publicly chosen minimum class-support threshold was used, and how were the fixed-label-set and support conditions maintained in every task? Please also specify the sensitivity and noise calibration for TabICL, TabPFN-v2, CatBoost, TabM, and the summarization baselines. These details could change the assessment of the privacy–utility comparison. (Sections 3.1, 3.5, and 4.1; Appendix D.1.)

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3kzh (1/2)

We thank the reviewer for the careful and constructive assessment. We address each point below and all changes are marked in blue in the revised PDF.

W1. Contribution of the quantum predictor, hardware utility, and cost:

(1) Head-level significance. We have now tested the head-level contrast directly and add the results in Table 4 (Appendix D.2). Paired Wilcoxon signed-rank tests across 200 TALENT datasets, with Holm correction, compare QICL against MLP, Fourier, and Transformer heads under the identical privatized interface. The Holm-adjusted pp-values are 4.13×10−54.13\times10^{-5}, 4.91×10−74.91\times10^{-7}, and 4.37×10−74.37\times10^{-7}, respectively. The advantage is modest but statistically detectable and consistent, not attributable to the interface alone.

(2) Hardware utility. We agree the original real-device experiment established executability, not practical utility. The model was trained in noiseless simulation and executed directly on hardware, which explains the large gap. We now perform noise-aware training: we download the ibm_kingston noise model, train locally under it with 6464 shots, and evaluate on the device with the same nine datasets, shots, and privacy budget. Hardware accuracy improves to 0.7080.708, versus 0.7510.751 in noiseless simulation. Since privacy noise is identical across settings, the residual gap reflects quantum execution effects: finite-shot estimation, gate and readout errors, and noise-model/device mismatch at execution time. We have updated Section 4.2 to include the new results.

(3) Latency and execution cost. For the nine-dataset evaluation we submitted 27,68927{,}689 circuits at 6464 shots each, totaling 1,772,0961{,}772{,}096 measurement shots. Aggregate QPU execution time was 99 minutes 1818 seconds; average queue waiting time per dataset-level submission was approximately 3.413.41 minutes, varying with shared-QPU load. These new experimental results and the corresponding analysis have been incorporated into Section 4.2 and Appendix D.3 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 3kzh (2/2)

W2 / Question. Auditability of the privacy comparison:

We thank the reviewer for highlighting the need for a more transparent and auditable privacy comparison. All methods are evaluated under the same record-level replace-one adjacency for the protected labeled context and the same target (ϵ,δ)(\epsilon,\delta)-DP guarantee.

(1) Fixed label set. This condition is enforced by construction, not by the classes observed in the private context. Each TALENT task provides number-of-classes metadata (n_classes or num_classes), defining a predetermined candidate label set Y={0,…,C−1}\mathcal{Y}=\{0,\ldots,C-1\}. The client builds CC class slots in fixed order, including empty ones, so replacing a context record may change class counts but cannot alter the label set or the dimensionality of the uploaded interface.

(2) Implementation of nmin⁡n_{\min}. In our implementation, when reliable public information about class support is unavailable, we select a practical denominator floor from the already privatized class priors. Specifically, with πc=nc/Nctx\pi_c=n_c/N_{\mathrm{ctx}}, we compute

n~c=Nctxπ~c,n^min⁡=max⁡(n0, ⌊[min⁡c∈Yn~c]+⌋),\tilde n_c=N_{\text{ctx}}\tilde\pi_c,\qquad \hat n_{\min}=\max\big(n_0,\ \lfloor[\min_{c\in\mathcal Y}\tilde n_c]_+\rfloor\big),

where [x]+=max⁡(0,x)[x]_+=\max(0,x) and n0=2n_0=2 is a predefined public lower bound. We use the clamped summary Sc(n^min⁡)=∑i:yi=cψper(hi)max⁡(nc,n^min⁡)S_c^{(\hat n_{\min})}=\frac{\sum_{i:y_i=c}\psi_{\text{per}}(h_i)}{\max(n_c,\hat n_{\min})}. Since ∥ψper(hi)∥2≤1\lVert\psi_{\text{per}}(h_i)\rVert_2\le1, the global sensitivity satisfies ΔS(n^min⁡)≤22n^min⁡\Delta_S(\hat n_{\min})\le\frac{2\sqrt{2}}{\hat n_{\min}} for every value of n^min⁡≥1\hat n_{\min}\ge1, even when the actual class support is smaller than n^min⁡\hat n_{\min}. Because n^min⁡\hat n_{\min} depends on π~\tilde\pi, the release is sequential: the client releases π~\tilde\pi, computes n^min⁡\hat n_{\min}, and then releases the summary with noise calibrated to ΔS(n^min⁡)\Delta_S(\hat n_{\min}). By adaptive composition of Gaussian mechanisms, this sequential release satisfies the same (ϵ,δ)(\epsilon,\delta)-DP as Theorem 1, so the threshold selection incurs no additional privacy cost (Appendix D.1). Note that this rule is only one possible implementation: when public, non-sensitive background knowledge about class sizes is available, it can be used to obtain a more accurate estimate of nmin⁡n_{\min} and further improve the privacy–utility trade-off, without affecting the privacy guarantee.

(3) Baseline sensitivity and noise calibration. All baselines use the same record-level replace-one adjacency, and block-wise Gaussian mechanism as QICL, and protect both features and labels. For each method mm and dataset jj, the clipping bound Bm,jB_{m,j} is the 90th percentile of representation norms on a held-out validation split excluded from the protected context: Bm,j=Q0.9({∥zi(m)∥2:i∈Djval})B_{m,j}=Q_{0.9}\big(\lbrace\lVert z_i^{(m)}\rVert_2: i\in D_j^{\text{val}}\rbrace\big), fixed before perturbation. Each representation is clipped as zˉi(m)=zi(m)min⁡(1,Bm,j/∥zi(m)∥2)\bar z_i^{(m)}=z_i^{(m)}\min\big(1,B_{m,j}/\lVert z_i^{(m)}\rVert_2\big).

  • CatBoost, TabM: perturbation on clipped input records before fitting;
  • TabICL, TabPFN-v2: perturbation on bounded record-local representations before context-dependent interaction;
  • TabICL-Mean: class-wise aggregation over clipped embeddings, so same analysis as QICL;
  • TabICL-AttnPool, TabICL-SetTrans: one record affects multiple attention weights, so mean-pooling bounds do not apply; perturbation applied before attention aggregation.

For the record-level baselines, each record is released as a feature block with sensitivity ΔF(m,j)≤2Bm,j\Delta_F^{(m,j)}\le2B_{m,j} and a one-hot label block with sensitivity ΔY=2\Delta_Y=\sqrt2, perturbed independently across records with \begin{equation*} \sigma_F^{(m,j)}=\frac{k_{\epsilon,\delta},\Delta_F^{(m,j)}}{\zeta_F},\qquad \sigma_Y=\frac{k_{\epsilon,\delta},\Delta_Y}{\zeta_Y},\qquad k_{\epsilon,\delta}=\frac{\sqrt{2\ln(1.25/\delta)}}{\epsilon}, \end{equation*} where the value of ζF\zeta_F and ζY\zeta_Y match the allocation (ζS,ζπ)(\zeta_S,\zeta_\pi) used in QICL (ζY=0.5\zeta_Y=0.5 and ζF2+ζY2=1\zeta_F ^2 + \zeta_Y^2= 1). Privatized labels are decoded by arg⁡max⁡\arg\max (post-processing). Each release therefore satisfies (ϵ,δ)(\epsilon,\delta)-DP by the argument of Theorem 1. For ϵ∈{1,2,4}\epsilon\in\{1,2,4\}, the classical calibration also satisfies the analytic Gaussian condition (Appendix D.1). Identical privacy budgets thus yield identical formal guarantees under the same adjacency and privacy unit, while the noise magnitudes differ according to each method's sensitivity and released information. Full details are given in Appendix D.1.

AuthorsOpenReview ↗

Follow-up on Reviewer Feedback

In response to your comments, we have added statistically validated comparisons with classical prediction heads, improved hardware accuracy to 0.708 through noise-aware training, reported hardware execution costs, and provided detailed explanations of the privacy calibration and baseline guarantees.

Thank you again for your valuable suggestions, which have helped us strengthen both the theoretical analysis and experimental evaluation.

We would appreciate it if you could reconsider your rating in light of these additional results and clarifications, and we would be happy to address any remaining concerns.

Reviewer_Fa5aOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a quantum-classical framework designed to protect the privacy of data sent to a server. It introduces a privacy-aware split quantum-classical framework (QICL) that divides processing between the client and a remote server. On the client side, row-level contextual representations are converted into bounded, class-wise summaries, with privacy protected by a block-wise Gaussian mechanism. On the server side, a candidate-aware relational predictor models interactions among each candidate and its strongest competitors before quantum inference.

The paper’s main contributions are a unified quantum-classical framework for tabular in-context learning that combines privacy-preserving task summaries with communication-efficient, resource-constrained quantum inference. It introduces a differentially private structured summarization module and a candidate-aware quantum predictor that emphasizes discriminative relationships between classes.

The validation results on benchmark datasets show a favorable privacy–utility–communication trade-off compared with classical baselines under matched settings.

Critical strengths

The main strength of the paper is the idea of a candidate-aware quantum predictor.

Critical weaknesses
  1. In the Background and Literature Review, the authors discuss classical in-context learning for tabular data and note that this area remains unexplored in quantum machine learning. Further, the authors apply differential privacy to the summary data sent to the server, a technique that has been extensively studied in the literature. This integrative approach has minimal novelty, and the contribution is incremental.
  2. The authors state that the server is untrusted. This raises questions about having the server perform the prediction task through a forward pass, since it cannot be trusted. It is understandable that data in transit requires privacy. However, with an untrusted server, the authors need to clarify the problem setting more clearly.
  3. The paper proposes preserving privacy using differential privacy, in which Gaussian noise is added to the summary data. In general, the main challenge for quantum computers is hardware noise. The authors did not discuss how the noise for privacy would affect prediction accuracy.
  4. The paper needs to explicitly discuss the limitations.
Questions for authors
  1. Can the authors define in detail what an “untrusted server” means in the context of the paper?
  2. Could the authors explain how privacy-related data noise affects the quantum computer’s predictions?
8 comments
AuthorsOpenReview ↗

Response to Reviewer Fa5a (1/2)

We thank the reviewer for the constructive comments and for recognizing the candidate-aware quantum predictor. We address each point below and all changes are marked in blue in the revised PDF.

W1. Novelty: We agree that tabular ICL and the Gaussian mechanism are established, and we claim novelty in neither. Our contribution is the interface that makes them work together under joint privacy, communication, and qubit constraints—a setting not previously addressed. Concretely:

(1) Jointly designed structured interface. The summary format and quantum encoding are co-designed for privacy sensitivity and qubit/communication budget, rather than applying DP post hoc; this coupling is what keeps the privacy guarantee meaningful under encoding.

(2) Formal guarantees under joint constraints. We provide a privacy guarantee (Theorem 1) and utility bound (Proposition 2) for the composed pipeline. A new remark (Remark 1) uses this condition to show that record-level release accumulates perturbation that grows with NctxN_{\mathrm{ctx}}, whereas the class-wise interface keeps it bounded by the number of classes (Section 3.5).

(3) Empirical validation. The interface outperforms record-level privatization of tabular ICL representations and alternative summarizers paired with the same quantum predictor (Section 4).

We have revised the introduction and related work to frame the contribution as this integrative interface, and we now state explicitly which components are borrowed versus novel (Section 1 and Section 2).

W2 / Q1: Meaning of untrusted server: We thank the reviewer for highlighting the need to clarify the trust assumptions. Our setting follows the widely adopted honest-but-curious (semi-honest) server model in privacy-preserving distributed and split learning [1, 2]. The server follows the prescribed inference protocol and executes the forward pass correctly, but may attempt to infer information about individual context records from what it receives. Prior work shows that even an honest-but-curious split-learning server can reconstruct private client information from intermediate representations, motivating explicit protection before transmission.

The client retains the raw labelled context locally and applies the Gaussian mechanism to the structured interface before uploading it. The server receives the privatized summaries and class priors, together with the separately encoded query. Under the stated adjacency relation, the DP guarantee protects the labelled context records. By the post-processing property of DP, any computation the server performs on F~\tilde F (rival selection, the concentration statistic, the quantum forward pass) cannot weaken the guarantee. The server can therefore compute on the interface without being trusted with the underlying records. A malicious server that returns incorrect predictions (integrity) is out of scope.

The client delegates prediction because quantum hardware is typically accessed through remote cloud providers [3]. For tractability, we have replaced "untrusted" with "honest-but-curious" throughout, added a threat-model in Section 3.1 listing what is observed and protected, and moved the malicious-server setting to the future work.

AuthorsOpenReview ↗

Response to Reviewer Fa5a (2/2)

W3 / Q2. Effect of privacy noise and quantum noise on predictions: Privacy noise and hardware noise enter at different stages and affect prediction differently. To address this question, we added an analysis of quantum noise in Appendix C.4. Appendix D.4 now includes the corresponding quantum-noise simulation. Moreover, Section 4.2 includes the updated hardware evaluation, which uses noise-aware training with the downloaded ibm_kingston noise model (further hardware results in Appendix D.3).

(1) DP noise. Added client-side before transmission and encoding. Proposition 2 gives a margin-based stability condition: under stated Lipschitz and rival-ranking assumptions, predictions are unchanged with probability ≥1−η\geq 1-\eta when the clean margin exceeds a perturbation-dependent bound set by the noise scales and interface dimensions. Empirically, Table 1 (200 TALENT datasets, ϵ∈{1,2,4}\epsilon \in \{1,2,4\}, δ=10−5\delta=10^{-5}) shows the expected privacy--utility trade-off.

(2) Quantum noise. Arises server-side during circuit execution and measurement. We consider depolarizing noise and finite-shot measurement noise, both of which constitute post-processing and cannot weaken the client-side DP guarantee. Under a binary Pauli-measurement model with depolarization probability pp and NshotN_{\mathrm{shot}} independent measurement shots, if the uploaded privatized interface satisfies (ϵ,δ)(\epsilon,\delta)-DP, the released binary-measurement outputs satisfy (ϵq,δq)(\epsilon_q,\delta_q)-DP, with

\begin{equation*} \epsilon_q =\log\left( \frac{ e^\epsilon+ \left(\frac{p}{2-p}\right)^{N_{\mathrm{shot}}} }{ 1+e^\epsilon \left(\frac{p}{2-p}\right)^{N_{\mathrm{shot}}} } \right), \qquad \delta_q = \frac{ 1-\left(\frac{p}{2-p}\right)^{2N_{\mathrm{shot}}} }{ 1+e^\epsilon \left(\frac{p}{2-p}\right)^{N_{\mathrm{shot}}} }\delta. \end{equation*} Consequently, ϵq≤ϵ\epsilon_q\leq\epsilon and δq≤δ\delta_q\leq\delta, indicating additional output-level privacy amplification under quantum noise model. This protects only an output-only adversary. The honest-but-curious server remains covered by the upload-level guarantee, and the amplification is limited in practice. Appendix C.4 further extends this bound from a single binary Pauli observable to the full predictor, which measures all nqn_q qubits over multiple candidate circuits. Empirically, at fixed ϵq=1\epsilon_q=1, lower depolarizing probability and more shots improve accuracy. QICL therefore relies on client-side Gaussian perturbation, not quantum noise, for privacy.

(3) Noise-aware hardware training. The original noiseless-training-then-hardware-execution protocol caused the large gap. We now train under the ibm_kingston noise model (6464 shots) and evaluate on-device with identical datasets, shots, and privacy budget: accuracy rises to 0.7080.708, versus 0.7510.751 in noiseless simulation. With DP noise identical across settings, the residual 0.0430.043 gap is purely quantum execution effects (finite shots, gate/readout errors, calibration drift, transpilation). Hardware noise is substantial without mitigation but largely recoverable via noise-aware training.

W4. Limitations: We agree that limitations should be more prominent and have expanded the discussion (Section 5) in the revised manuscript along three axes.

(1) Threat model. QICL assumes an honest-but-curious server that follows the prescribed inference protocol; actively malicious servers (e.g., returning incorrect predictions) are out of scope.

(2) Hardware. Hardware quantum noise, finite measurement shots, and limited real-device evaluation remain obstacles to practical deployment. Future work will develop hardware-aware training, error mitigation, and quantum hardware-in-the-loop calibration to improve reliability on physical devices.

(3) Privacy calibration. Our analysis relies on a minimum class-support threshold nmin⁡n_{\min}, while the implementation selects a practical denominator floor n^min⁡\hat n_{\min} from privatized class priors, which enforces the sensitivity bound by construction without additional privacy cost. However, selecting a single threshold based on the least-supported class may result in conservative noise calibration for well-populated classes. Future work could adopt individual privacy accounting to provide tighter guarantees for records in well-populated classes.

References

[1] Deng, Lu, and Duan. InfoDecom: Decomposing information for defending against privacy leakage in split inference. AAAI 2026.

[2] Wei et al. Federated learning with differential privacy: Algorithms and performance analysis. IEEE TIFS, 2020.

[3] Nguyen et al. Quantum cloud computing: A review, open problems, and future directions. arXiv:2404.11420, 2024.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Revisions

Based on your feedback, we have clarified the novelty of QICL and the honest-but-curious server assumption, distinguished privacy noise from quantum hardware noise, added further theoretical and experimental analyses, and expanded the discussion of limitations.

Thank you again for your constructive feedback, which has helped us improve the manuscript.

We would greatly appreciate it if you could review our revisions and consider updating your rating if you feel your concerns have been addressed, and we would be happy to answer any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.