acceptodds
Under review as a conference paper at ICLR 2027

Encryptability As a Coordinate Choice: Depth-One Homomorphic Federated Learning of Quantum Neural Networks

Abstract

Encrypted training relies on keeping server-side updates low-degree. This constraint traditionally excludes models whose weights inhabit a compact Lie group (notably variational quantum circuits, where every trainable weight is an rotation). Expressed in Euler angles or discrete alphabets, these updates appear transcendental, historically demanding prohibitive costs: one client–server round per gate, or upwards of operations per weight. This penalty is strictly an artefact of coordinates. In the unit-quaternion (spin) chart, group composition is exactly bilinear (degree two, with coefficients in ). Consequently, encrypted rotation updates cost one multiplicative level and federated averaging costs zero in any levelled homomorphic scheme, completely eliminating bootstrapping. This implementation-independent algebraic property is confirmed across two cryptographic backends, introducing only and  rad of aggregation error. Leveraging this reduction yields a non-interactive protocol for encrypted federated training of hybrid quantum–classical networks. It includes correctness proofs for aggregation and sign handling, plus a compilation lemma proving parameterised entanglers add only constant-factor overhead without altering the depth class. Empirically, a paired five-seed study confirms zero measurable utility tax ( MSE, ), and a noise-budget ablation falsifies the hypothesis that encryption noise regularises. These convergence trends replicate across datasets and scale to clients. Finally, hardware validation on a -qubit processor achieves fidelity against a unencrypted control.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
quantum machine learning, homomorphic encryption, federated learning, variational quantum circuits, lie groups, unit quaternions, privacy-preserving machine 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_HFnSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a quaternion-based homomorphic encryption framework for federated learning of quantum neural networks. By representing quantum rotations using unit quaternions, the method reduces the multiplicative depth of encrypted operations. The authors provide theoretical analysis and evaluate the approach using hybrid quantum-classical models and IBM quantum hardware. Results show similar performance between encrypted and plaintext training, although computational overhead remains substantial.

Critical strengths
  1. The paper presents an interesting approach to reducing encryption complexity through quaternion representations.
  2. The theoretical analysis is supported by experiments using different encryption backends, datasets, and real quantum hardware.
Critical weaknesses
  1. The main training experiments use Euclidean averaging rather than the proposed encrypted quaternion aggregation (Appendix G). The authors should evaluate the complete proposed protocol during training.
  2. The data partitioning strategy is unclear, and non-IID experiments are missing. The authors should evaluate how data heterogeneity affects aggregation and convergence.
  3. The hardware fidelity evaluation relies on Z-basis measurements (Appendix G.7), which cannot fully establish quantum state fidelity. More rigorous evaluation is needed.
  4. Comparisons with existing methods are insufficient. Direct comparisons with standard HE-FedAvg are needed to demonstrate the benefits of quaternion aggregation.
  5. Practical efficiency remains questionable. Encryption takes approximately 167 seconds per round, and encrypted CNOT operations take around 1.5 hours each. End-to-end runtime comparisons are needed.
  6. The experimental encryption parameters provide no formal security guarantee (Appendix J). Evaluation under realistic security settings is needed.
  7. Some presentation errors remain in Figure 3, including outdated editing references.
Questions for authors

see weeknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer HFnS

Thank you for an actionable review!

W1 — Euclidean averaging in training. Correct for the submitted curves, which came from the harness. E7 trains end-to-end with server-side homomorphic quaternion aggregation in every round (OpenFHE CKKS, 128-bit; five paired seeds, 5050 rounds). IID: Δ=−8.9×10−8\Delta = -8.9 \times 10^{-8} versus plaintext (95% CI [−4.2,+2.4]×10−7[-4.2, +2.4] \times 10^{-7}) at R2=0.66R^2 = 0.66 (Table 1).

W2 — partitioning and non-IID. E1/E5/E6 used the first 512512 California records in three contiguous client blocks, the next 128128 for validation. On that split, even ridge regression does not beat the training-mean predictor, so those runs test only the effect of encryption on the trajectory; Appendix G says so. E7 uses a random 2,048/1,0242{,}048/1{,}024 split, partitioned IID, geographically (kk-means on location; client sizes 175/1,152/721175/1{,}152/721) and by target tertile: every encrypted arm is equivalent to plaintext within ±0.01\pm 0.01 MSE (TOST p≤2.3×10−5p \le 2.3 \times 10^{-5}); label skew lowers R2R^2 to 0.360.36 for every arm alike. On Wine Quality (IID, red-vs-white, quality tertiles; R2=0.31/0.28/0.13R^2 = 0.31/0.28/0.13) the same holds (TOST p≤7.4×10−11p \le 7.4 \times 10^{-11}).

W3 — ZZ-basis fidelity. Agreed: it is an upper bound on state fidelity and blind to phase errors, and the submitted point was a single 6464-shot execution. We therefore ran an interleaved tomography experiment on ibm_fez (E11, Appendix G.14): unencrypted, padded and pad-only arms interleaved in one job on one qubit, barriers between stages, X/Y/ZX/Y/Z state tomography at 40964096 shots over 2020 angles, no mitigation. State fidelity: 0.98500.9850 (unencrypted) and 0.98120.9812 (padded); the pad costs 3.8×10−33.8 \times 10^{-3} (95% CI [0.8,6.7]×10−3[0.8, 6.7] \times 10^{-3}), inside the pre-specified ±0.01\pm 0.01 margin (TOST p=1.2×10−4p = 1.2 \times 10^{-4}). On the ZZ-basis statistic of the submission, the same arms score 0.99870.9987 and 0.99760.9976, which shows how much that statistic flatters.

W4 — comparison with standard HE-FedAvg. E7 has two depth-00 baselines: raw-angle HE-FedAvg (angles encrypted as reals, as in Dutta et al. 2024a) and circular HE-FedAvg on (cos⁡θ,sin⁡θ)(\cos\theta, \sin\theta). E7b and E7c then locate where the charts differ:

  • Branch cut (E7b; 1515 paired seeds, two initialisation widths). E7b isolates the condition under which angle averaging fails across the ±π\pm\pi cut. When clients upload unconstrained real-valued angles from a shared lift, Euclidean averaging avoids the cut (∣Δ∣≤3×10−5|\Delta| \le 3 \times 10^{-5}). When angles are confined to the principal branch (−π,π](-\pi, \pi] — as required by Ma & Li's mm-bit fraction-of-a-turn QHE encoding, by atan2\mathrm{atan2} gate recovery, or by bounded-domain polynomial approximations — clients straddle the cut in 27–29% of rounds, every straddling parameter is corrupted by 2π/3≈2.092\pi/3 \approx 2.09 rad (7878 and 6262 times), validation MSE jumps by up to 1.621.62 within one round, and final MSE degrades in 2727 of 3030 runs (mean +0.12+0.12 and +0.23+0.23; Wilcoxon p=0.002p = 0.002 and 6.1×10−46.1 \times 10^{-4}; TOST equivalence fails, p≥0.98p \ge 0.98). Hemisphere-aligned quaternion HE-FedAvg suffers zero corruptions.
  • General gates (E7c). For RXR_X-only circuits, the circular chart is as good as quaternions. When gates do not share an axis, a two-component chart cannot represent them: projecting trained SU(2)SU(2) gates onto one axis raises MSE from 0.310.31 to 0.870.87, while quaternion HE-FedAvg matches plaintext (Δ=−5.0×10−6\Delta = -5.0 \times 10^{-6}).
  • Delegation (E8b). The encrypted quaternion aggregate composes with an encrypted pad key in 11 level (7272 ms); from angles this needs 66 levels (coaxial) or 88 (Euler).
  • Bias. Quaternion averaging deviates from the geodesic mean by −(1/24)∑wδ3-(1/24)\sum w\delta^3, a quarter of the projected-matrix or full-angle value (E12: measured ratio 4.014.01).

So the benefit of the quaternion chart is not better accuracy on RXR_X circuits; it is that one representation serves general gates, principal-branch encodings and composition with pad keys at depth 11.

W5 — efficiency. The 167167 s per round and 22.6722.67 MB per client-round belong to the bit-level reference, which encrypts every fixed-point bit separately. In CKKS at 128-bit security, one 132132 kB ciphertext carries a client's whole model; per round, encryption takes 1.91.9 ms per client, aggregation 0.70.7 ms, decryption 0.80.8 ms. The ≈1.5\approx 1.5 h per encrypted CNOT concerns delegated execution in the bit-level reference; the federated loop evaluates none. Interactive baselines need 3737–7373 sequential exchanges per round.

W6 — security. All contexts of E7–E9 and E12 run under OpenFHE's HEStd_128_classic. An executable core-SVP audit (Appendix J) shows the bit-level profiles (n≤256n \le 256) are broken; no security is claimed for them.

W7 — Figure 3. Thank you. The figures are regenerated from the result files without editing labels.

We hope the revision addresses each point and would be grateful if you took it into account in your assessment.

Thanks for addressing those raised issues, and it clarifies some of my concerns.

Thanks a lot for reading our reply! Can we be of any further help? Happy to do extra experiments/provide extra clarifications.

Reviewer_6dbLOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that the homomorphic encryptability of group-valued parameters is a coordinate choice, not an architectural one. QNN weights are SU(2) rotations; in Euler angles the composition law is transcendental, in unit quaternions it is bilinear (Prop. 1), so an encrypted rotation update costs one multiplicative level and encrypted FedAvg zero, with no bootstrapping. On that basis it builds a non-interactive encrypted federated protocol over the quaternion QFHE of Ma and Li (2022), proves aggregation and sign-handling correctness (Lemma 1), derives the exact cubic discrepancy between quaternion chordal averaging and Euclidean angle averaging (Prop. 2), compiles parameterised entanglers at constant overhead (Lemma 2), and reports a two-backend implementation, a federated training study, and an ibm_fez round trip.

Critical strengths
  1. The depth ledger and the geometry result. Prop. 1 and Cor. 1 give a usable statement of what a levelled scheme can reach, and Prop. 2 is a correct, new, closed-form result with a clear practitioner reading.

  2. Implementation discipline. Two backends, paired seeds, a round-0 equality check and an 8-bit precision falsification ablation. The experimental hygiene is better than the field norm, and the paper is candid about what is calibrated rather than measured.

Critical weaknesses
  1. The headline mechanism is prior work, and the framing does not say so. The paper's Eq. (1) is Ma and Li's Eq. (2.14); Eq. (8) is their Definition 2.2, down to the same unusual (iX,iZ,iY)(iX, iZ, iY) assignment; the Pauli/quaternion track conversion is their Prop. 4.3 and Lemma 5.1; the encrypted conditional rotation is their Section 3.2. What Ma and Li do not do is quantify the depth of the key update - that, Prop. 2, the federated protocol and the first implementation are this paper's work. Section 3 acknowledges the proximity; the title and abstract do not.

  2. The mechanism is never exercised end to end. The learning curves come from a masked-aggregation harness, so no training run passes through the protocol's homomorphic aggregation; the delegation channel, where the quantum machinery is irreplaceable, is validated at two CNOTs. The ibm_fez test is one qubit and one rotation, and its 0.0078 fidelity gap is not explained by the paper's own decomposition: the simulator (0.9989) and the matched unencrypted control (0.99957) together predict F≈0.9985F \approx 0.9985, leaving about 6.7×10−36.7 \times 10^{-3} of infidelity unaccounted for. Readout error is common to both arms and cancels in their difference, so it cannot explain the gap.

Questions for authors
  1. What rotation axes do the 36 trainable angles use, and is the recovery formula generalised accordingly? Which of Eq. (2) and Eq. (8) has the y/zy/z assignment right?

  2. What is the multiplicative depth actually required by the underlying MHE scheme in the reported configurations, how does it relate to 3kL+13kL+1, and how does it fit inside q=256q=256 and q=232q=2^{32}? What are the corrected CKKS parameters, and what do the aggregation numbers look like under them?

  3. For the hardware run: how is the analytic reference computed, is any post-processing applied to the counts, and were the encrypted and control shots interleaved? What accounts for the 6.7×10−36.7 \times 10^{-3}?

  4. Can you show one end-to-end federated run in which the server performs the protocol's homomorphic aggregation, with the total depth and wall-clock cost of the rotation path including track switching?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6dbL (Part 1/2)

Thank you for a precise review, and for identifying Proposition 2 as "a correct, new, closed-form result".

W1 — attribution. Your mapping is exact: our Eq. 1 is Ma & Li's Eq. 2.14; Eq. 8 is their Definition 2.2, including the (iX,iZ,iY)(iX, iZ, iY) basis; the track conversion is their Prop. 4.3 and Lemma 5.1; the encrypted conditional rotation is their §3.2. The abstract and introduction should have said so. The revised PDF credits them in the abstract, introduction, Figure 1, Related Work and §4.1, and lists as ours what you listed: the depth of the key update, Prop. 2, the federated protocol and the implementation. We propose the camera-ready title "Homomorphic Federated Learning of Quantum Neural Networks in the Quaternion Chart".

W2 / Q4 — end to end. Correct: the submitted learning curves came from the harness. E7 runs the protocol's homomorphic aggregation in every round:

  • Setup. OpenFHE CKKS (HEStd_128_classic, ring 2132^{13}), one ciphertext per client per round holding all 3636 quaternions and the 968968 classical parameters. The server adds nkn_k-fold copies (00 levels); clients decrypt, divide, renormalise. Five paired seeds, 5050 rounds, K=3K = 3, California Housing (random 2,048/1,0242{,}048/1{,}024 split).
  • IID. Plaintext 0.3236±0.03710.3236 \pm 0.0371 (R2=0.66R^2 = 0.66); quaternion HE-FedAvg Δ=−8.9×10−8\Delta = -8.9 \times 10^{-8} (95% CI [−4.2,+2.4]×10−7[-4.2, +2.4] \times 10^{-7}).
  • Non-IID. Geographic and label-tertile partitions: every encrypted arm is equivalent to plaintext within ±0.01\pm 0.01 (TOST p≤2.3×10−5p \le 2.3 \times 10^{-5}), with raw-angle and circular HE-FedAvg as baselines.
  • Simulator verification. The exact statevector simulation of E7 matches Qiskit's EstimatorQNN parameter-shift evaluation of the same circuit to 4.4×10−164.4 \times 10^{-16} in outputs and in all 1,0041{,}004 gradients (Appendix G.4).

Total depth and wall-clock of the rotation path.

  • Aggregation, every round: d=0d = 0, ℓ=0\ell = 0; 1.91.9 ms encryption per client, 0.70.7 ms aggregation, 0.80.8 ms decryption.
  • From the encrypted aggregate to the delegated key (E8b). The server composes the unnormalised encrypted aggregate with an encrypted pad key, without decrypting: 0+10 + 1 levels in total, ring 2142^{14}, 7272 ms for 3636 wires, key recovered to 3.4×10−73.4 \times 10^{-7} rad (8.7×10−98.7 \times 10^{-9} at 45-bit scaling).
  • LL private gates on the key: exactly ⌈log⁡2(L+1)⌉\lceil \log_2(L+1) \rceil levels (1,2,3,41, 2, 3, 4 for L=1,2,4,8L = 1, 2, 4, 8), 0.060.06–0.930.93 s. From Euler angles the same update takes 88–1111 levels at ring 2152^{15} and 1.51.5–18.718.7 s.

Track switching. The cost comes from here. In the bit-level reference, each encrypted CNOT layer advances the key chain by 3k3k levels (k=12k = 12, so 3636) and costs ≈1.5\approx 1.5 h per CNOT in the single-threaded Boolean reference. E7 never evaluates a CNOT homomorphically, and the delegation channel is validated only at E4's two CNOTs; both facts are stated as limitations. We have not implemented track switching in CKKS and make no measured claim about it.

AuthorsOpenReview ↗

Response to Reviewer 6dbL (Part 2/2)

Q1 — axes; Eq. 2 vs Eq. 8. In the submitted runs all 3636 trainable gates are RXR_X, encoded as t=(cos⁡(θ/2),−sin⁡(θ/2),0,0)t = (\cos(\theta/2), -\sin(\theta/2), 0, 0). You are right about the implication: such gates are coaxial, two components suffice, and the circular chart (cos⁡θ,sin⁡θ)(\cos\theta, \sin\theta) aggregates them equally well; the revised paper says so and no longer rests the case for four components on RXR_X circuits. E7c tests the case that needs them: every trainable gate is a general SU(2)SU(2) element. Quaternion HE-FedAvg then matches plaintext (Δ=−5.0×10−6\Delta = -5.0 \times 10^{-6}; largest per-seed gap 2.4×10−52.4 \times 10^{-5}; per-round aggregation error ≤1.9×10−7\le 1.9 \times 10^{-7} rad), 174174 of 180180 trained gates have all four components above 10−310^{-3}, and projecting the gates onto one axis raises MSE from 0.310.31 to 0.870.87. The general ansatz is not more accurate than RXR_X on this task (p=0.28p = 0.28); the point is representability.

Eq. 8, Ma & Li's (iX,iZ,iY)(iX, iZ, iY), is the convention for which Φ(q⊗q′)=Φ(q)Φ(q′)\Phi(q \otimes q') = \Phi(q)\Phi(q'). Eq. 2 as printed used (X,Y,Z)(X, Y, Z) order, which is anti-homomorphic; the two coincide for RXR_X. The revised Eq. 2 uses Ma & Li's basis, q(θ,u^)=(cos⁡(θ/2),−uxsin⁡(θ/2),−uzsin⁡(θ/2),−uysin⁡(θ/2))q(\theta, \hat{u}) = (\cos(\theta/2), -u_x \sin(\theta/2), -u_z \sin(\theta/2), -u_y \sin(\theta/2)), with recovery θ^=2⋅atan2(−⟨q1:3,u^′⟩,q0)\hat{\theta} = 2\cdot\mathrm{atan2}(-\langle q_{1:3}, \hat{u}' \rangle, q_0), u^′=(ux,uz,uy)\hat{u}' = (u_x, u_z, u_y). Composition against 2×22 \times 2 unitaries is checked to 2.5×10−162.5 \times 10^{-16}.

Q2 — MHE depth, 3kL+13kL + 1, CKKS parameters. 3kL+13kL + 1 counts key pairs in Ma & Li's key chain, not multiplicative depth: each encrypted CNOT layer advances the chain by 3k3k levels (3737 key pairs for k=12k = 12, L=1L = 1). Moving along the chain is level promotion: a ciphertext is homomorphically decrypted under the next key (Mahadev's dual-scheme conversion), a bootstrapping-type refresh along an acyclic chain; that is why q=256q = 256 suffices structurally. "No bootstrapping" is restricted to the federated loop in CKKS, and bit-level aggregation is a ripple-carry adder, exact but not depth 00.

Security of q=256q = 256 and q=232q = 2^{32}: none. Appendix J has an executable core-SVP audit (primal and dual, validated on NewHope/JarJar/BCNS): n=1n = 1 falls to exhaustive search, n=64n = 64 and 256256 are broken at the smallest block size covered; with the scheme's own noise rule, n=1024n = 1024 reaches about 115115 bits and n=2048n = 2048 about 294294.

CKKS: for the E2 configuration (depth 4040, 50-bit scaling, HEStd_128_classic) OpenFHE 1.6.1 selects ring 2172^{17}, not the 2162^{16} stated in the submission; Table 4 and Appendix J state 2172^{17}. The aggregation path needs neither: depth 00, ring 2132^{13}; distinct client vectors (K∈{3,20}K \in \{3, 20\}) aggregate within 7.7×10−87.7 \times 10^{-8} rad, and Prop. 2 matches the realised gap to three digits.

Q3 — hardware. You are right on each point.

  • The submitted comparison. The encrypted point (FZ=0.9918F_Z = 0.9918) came from a single 6464-shot execution at θ=1.721\theta = 1.721 rad (3333 zeros; classical fidelity against the analytic P(0)=cos⁡2(θ/2)P(0) = \cos^2(\theta/2), no post-processing or mitigation), compiled at optimisation level 33 without barriers between pad, gate and unpad, so the stages could be fused; the unencrypted control was a separate, non-interleaved run at 40964096 shots. At 6464 shots an ideal noiseless channel scores 0.9960±0.00560.9960 \pm 0.0056 (P(FZ≤0.9918)=0.13P(F_Z \le 0.9918) = 0.13), so the 6.7×10−36.7 \times 10^{-3} gap is consistent with shot noise alone. Readout error cancels between matched arms, as you say, and the ZZ-basis statistic is an upper bound on state fidelity, not a lower bound. Simulated round trips of the same pipeline (1010 angles ×\times 10241024 shots) give 0.999850.99985, at the finite-shot expectation of 0.999760.99976.
  • What replaces it: E11 on ibm_fez (Appendix G.14). Because runs with mismatched shot budgets and ZZ-basis-only measurement cannot isolate the physical overhead of the pad, we ran an interleaved experiment: 2020 angles ×\times 33 interleaved arms (unencrypted, padded, pad-only) ×\times X/Y/ZX/Y/Z state tomography at 40964096 shots on physical qubit 114114 (chosen from the live calibration), with stage barriers preventing gate fusion, no readout mitigation, and a ±0.01\pm 0.01 equivalence margin fixed in advance.
  • Result. State fidelities are 0.98500.9850 (unencrypted), 0.98120.9812 (padded) and 0.99730.9973 (pad-only): carrying the quaternion pad through the rotation costs 3.8×10−33.8 \times 10^{-3} of state fidelity (95% CI [0.8,6.7]×10−3[0.8, 6.7] \times 10^{-3}; paired-tt p=0.014p = 0.014), well inside the margin (TOST p=1.2×10−4p = 1.2 \times 10^{-4}). The 1.51.5–1.9%1.9\% infidelity of the rotated arms is consistent with unmitigated relaxation-type readout asymmetry (fitted P(0∣1)≈2.3%P(0|1) \approx 2.3\% vs 0.1%0.1\% calibrated), which both arms share. The CKKS Case A key update (11 level, ring 2142^{14}) reproduces the unpad unitaries executed on ibm_fez to 1.5×10−101.5 \times 10^{-10} in operator norm.
Reviewer_aDWaOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a quaternion-based approach to encrypted federated learning for variational quantum neural networks. Its central observation is that the composition of SU(2) rotations becomes a degree-two polynomial operation when represented using unit quaternions, whereas the corresponding operation in other parameterizations may involve trigonometric functions or require gate decompositions.

The paper develops a proposed protocol combining quaternion-based rotation representations, homomorphic encryption of quaternion components, and Pauli one-time-pad-based quantum-state encryption for delegated quantum computation. The authors argue that rotation updates can be evaluated with at most one multiplicative level, while weighted federated averaging requires only ciphertext addition and plaintext scaling. They further present correctness arguments for aggregation, quaternion sign and branch handling, and the compilation of parameterized entanglers.

The empirical evaluation includes a six-qubit, six-layer hybrid quantum–classical model on California Housing and Wine Quality, comparisons between fixed-point multi-key encryption and CKKS, paired-seed experiments, aggregation-geometry tests, scaling experiments involving up to 20 clients, and single-qubit rotation experiments on IBM quantum hardware. The reported results suggest negligible numerical discrepancies in the tested aggregation operations and similar predictive losses between plaintext and encrypted configurations.

The paper addresses a potentially interesting intersection of quantum machine learning, homomorphic encryption, and federated learning. However, the central claims require more rigorous validation of the relationship between the proposed algebraic reduction, the actual encrypted computation, the security guarantees, and the end-to-end federated-learning protocol. In particular, the manuscript does not yet establish that its depth-one analysis translates into the claimed secure and efficient training protocol under a consistent threat model.

Specifically, compared with existing studies, the actual incremental contribution of this work is not as substantial as implied in the article. The work of Ma and Li (2022) cited in the paper has already applied quaternion representation to single-client delegated quantum computing, meaning that this core algebraic result is not original to this paper. The innovations of this paper lie in federated wrapper, the correctness & geometry analysis of aggregating group-valued updates, and the claimed working implementation of continuous-rotation QHE. However, the authors reveal in the discussion chapter that the implementation overhead in the delegated execution branch remains approximately 1.5 hours per encrypted CNOT, rendering it not yet practical for deployment. Accordingly, the core claim of the paper is not clearly stated: although the bottleneck has been eliminated at the algebraic level, it has not been resolved at the engineering operation level.

The main depth-reduction claim is not sufficiently established for the private-parameter setting, the stated shared-key model does not provide the claimed confidentiality between clients, and the experiments do not validate the full security and efficiency claims of the proposed protocol. These foundational concerns affect the interpretation of the contribution, rather than being issues that can be resolved by improving presentation or adding routine experiments. I would reconsider the rating if the authors can demonstrate that the protocol performs the claimed operations on the intended encrypted operands, provide a consistent security argument under a clearly specified threat model, and substantiate the resulting efficiency and novelty relative to existing quantum homomorphic-encryption methods.

Based on the current paper content, I do not consider it suitable for publication at ICLR, so I have given a clear rejection decision. Nevertheless, it is at least a complete and reasonable submission. I have not observed obvious signs of AI generation in the text, and its quality is markedly higher than some AI research papers.

Critical strengths

The experimental design of this study is relatively rigorous.

The paper reports experiments using two homomorphic-encryption backends, paired-seed comparisons, aggregation-error measurements, communication-cost estimates, and a hardware-based single-qubit rotation experiment. It also acknowledges several limitations, including substantial bandwidth requirements, expensive encrypted entangler evaluation, and the absence of robustness against malicious clients.

The experimental design has several commendable aspects. The five-seed paired comparison explicitly checks pairing consistency: the plaintext and encrypted arms have identical round-0 losses, and the authors report a 95% confidence interval for the mean paired difference alongside both a paired t-test and a Wilcoxon signed-rank test. The 8-bit precision experiment further investigates the hypothesis that encryption noise acts as a beneficial regularizer. The noisier configuration shows no statistically significant improvement, providing evidence against this explanation, although it does not conclusively rule it out.

Finally, Section 5.1 explicitly distinguishes mechanism-level claims from run-level empirical claims. This distinction is methodologically appropriate, although the extent to which the broader protocol-level conclusions are supported still depends on the correctness, security, and scope of the underlying experiments.

Critical weaknesses
  1. The central multiplicative-depth claim does not adequately distinguish plaintext–ciphertext computation from ciphertext–ciphertext computation, undermining the claimed contribution.

The main technical argument is that quaternion composition is bilinear and therefore can be evaluated at multiplicative depth one. However, Proposition 1(ii) explicitly assumes that the quaternion �0��0� representing the applied rotation is plaintext, while �1��1� is encrypted. Under this assumption, every component of �2��2� is a linear combination of the encrypted components of �3��3� with plaintext coefficients.

This distinction is fundamental. In standard homomorphic-encryption accounting, multiplication by a plaintext scalar or plaintext constant does not ordinarily incur ciphertext–ciphertext multiplicative depth. Some approximate-arithmetic schemes may consume a modulus level or incur rescaling costs, but these are not equivalent to ciphertext–ciphertext multiplicative depth. The paper needs to define its cost model precisely and explain why the claimed depth-one bound is meaningful under that model.

More importantly, if both �4��4� and �5��5� are encrypted because the rotation parameters themselves must remain confidential, their composition generally requires encrypted–encrypted multiplication. Proposition 1(ii) does not establish the claimed cost for this case. Conversely, if �6��6� is public to the server, the construction does not hide that rotation parameter from the server. The manuscript does not adequately reconcile these alternatives with its intended privacy guarantees.

This issue is central rather than cosmetic: the paper's main efficiency claim depends on what is encrypted, which party knows each rotation, and which operations are actually evaluated homomorphically.

The novelty argument also requires sharper positioning against Ma and Li, Quantum Fully Homomorphic Encryption by Integrating Pauli One-time Pad with Quaternions (Quantum, 2022). That work already uses quaternion representations for single-qubit quantum homomorphic evaluation and develops a conversion between quaternion and Pauli encryption schemes.

The present manuscript may offer a useful federated-learning application or a different aggregation analysis, but it must isolate the genuinely new construction and establish its complexity advantage over the existing scheme. As currently presented, the algebraic observation alone is insufficient to support the broad claim of a new depth-one encrypted-training paradigm.

  1. The security model, protocol description, and empirical validation do not establish the claimed end-to-end secure federated-learning system.

The paper combines two distinct cryptographic mechanisms: homomorphic encryption of classical quaternion components for model-update aggregation and quantum-state encryption using quaternion and Pauli one-time pads for delegated quantum computation. Their security properties are not interchangeable. The fact that a quantum state becomes maximally mixed when averaged over a suitable random pad does not, by itself, establish confidentiality of classical model updates, security among clients, or security of the complete composed protocol.

A particularly important inconsistency concerns the shared-key assumption. The manuscript states that clients share homomorphic decryption capability and that the server cannot decrypt the ciphertexts. However, Appendix J also acknowledges that a peer client holding another client's ciphertext can decrypt it under this shared-key model. Consequently, the statement that peers learn only the decrypted global aggregate is not a cryptographic guarantee under the stated threat model: a curious client with access to an individual ciphertext and the shared decryption capability can inspect that ciphertext directly. This is especially consequential because protecting client updates from other participants is one of the motivations for federated confidentiality.

The protocol's treatment of server-side rotation updates also remains unclear. The manuscript describes clients training locally in plaintext, encrypting the resulting quaternion components, and the server optionally applying a plaintext rotation before aggregation. It does not sufficiently specify how these operations correspond to the actual federated optimization update, which party possesses the rotation parameters, and whether those parameters are intended to remain private. A precise end-to-end protocol and a security argument covering all transmitted objects are necessary to substantiate the claimed confidentiality.

The empirical evidence does not resolve these concerns. The paired-seed experiment evaluates a small six-qubit hybrid model on two tabular datasets, while the hardware experiment evaluates single-qubit rotation behavior. These experiments provide some evidence about numerical consistency under the tested configurations, but they do not independently demonstrate the security of the complete protocol or the claimed end-to-end efficiency of encrypted federated training. In particular, near-identical plaintext and encrypted losses show that the tested encryption/decryption path introduces little measurable numerical disturbance; they do not establish that the proposed protocol delivers the claimed security and computational advantages over appropriate baselines.

The paper also reports a structural-verification cryptographic parameter profile that explicitly carries no formal security guarantee for the reported runs. This is a legitimate way to test an implementation, but it cannot substantiate an empirical claim of secure deployment. In combination with the unresolved shared-key issue and the ambiguity about which operations are performed on encrypted data, this substantially limits the conclusions supported by the experiments.

Taken together, these issues prevent me from verifying the paper's central claim that the proposed method simultaneously achieves the advertised multiplicative-depth reduction, confidentiality, and practical end-to-end federated training.

Questions for authors
  1. What is the precise encrypted computation underlying Proposition 1(ii)?

The proposition assumes plaintext �0��0� and encrypted �1��1�. Can the authors specify which party knows �2��2�, whether it is a public gate parameter or a private trainable parameter, and what operation is performed when both operands must be encrypted? Please provide the multiplicative-depth and modulus-level accounting for each case, distinguishing plaintext–ciphertext from ciphertext–ciphertext multiplication. In particular, why should the plaintext–ciphertext linear map be counted as consuming one multiplicative level under the stated homomorphic-encryption model?

  1. How is confidentiality between clients guaranteed under the shared-key assumption?

Appendix J acknowledges that a peer holding another client's ciphertext can decrypt it. Does the intended threat model exclude clients that inspect other clients' ciphertexts, or is there an additional cryptographic mechanism that prevents this? Please state explicitly which parties can decrypt individual updates, which can decrypt only the aggregate, and what confidentiality guarantee holds against a curious participating client.

  1. What is the exact end-to-end protocol being evaluated, and how does it differ from Ma and Li (2022)?

Please identify the operations that are new relative to the existing quaternion-based quantum homomorphic-encryption scheme, including the precise role of federated averaging, encrypted parameter updates, and conversions between the quaternion and Pauli tracks. Which operations in the reported experiments are actually performed homomorphically, and which are simulated or executed in plaintext? Please enhance the relevant expressions in the original text. This clarification is necessary to assess both novelty and whether the experiments validate the central claim.

  1. What security guarantees apply to the experimental configuration, as opposed to the abstract scheme?

The paper explicitly labels the default cryptographic profile as suitable for structural verification rather than as a secure parameterization. Can the authors distinguish the properties proved for the abstract construction from those demonstrated by the actual implementation, and specify which reported experiments use a parameter set with a defensible security level? How should the reported results be interpreted when the tested profile carries no formal security guarantee?

4 comments
AuthorsOpenReview ↗

Response to Reviewer aDWa (Part 1/2)

Thank you for a careful review!

W1 / Q1 — depth accounting. You are right that Prop. 1(ii) assumed a plaintext gate rr and charged a multiplicative level to what is a plaintext–ciphertext linear map; its ciphertext–ciphertext depth is 00. The revised Prop. 1 is a three-case ledger with two counters, ciphertext–ciphertext depth dd and CKKS rescale levels ℓ\ell:

  • (A) public gate rr ×\times encrypted key tt. (t⊗r)i=∑jcijtj(t \otimes r) _ i = \sum _ j c _ {ij} t _ j with plaintext cijc _ {ij}: 1616 plaintext–ciphertext products, d=0d = 0, ℓ=1\ell = 1 (one rescale, since cijc _ {ij} is encoded at scale Δ\Delta). The processor knows rr.
  • (B) private gate ×\times encrypted key. Trainable weight supplied encrypted: 1616 ciphertext–ciphertext products with relinearisation, d=1d = 1, ℓ=1\ell = 1; LL gates compose at d=⌈log⁡2(L+1)⌉d = \lceil \log_2(L+1) \rceil by a balanced tree. This is the both-encrypted case you asked about; the processor learns neither operand.
  • (C) FedAvg. ∑knk⋅Enc(q(k))\sum_k n_k \cdot \mathrm{Enc}(q^{(k)}) by additions only: d=0d = 0, ℓ=0\ell = 0 with integer weights (clients divide by nn after decryption); ℓ=1\ell = 1 with fractional weights.

Measured on the intended encrypted operands (OpenFHE CKKS, 128-bit, 3636 gates per ciphertext; E8, E8b):

  • (A) 11 level, 1212 ms; (B) 11 level, 1616 ciphertext products, 5252 ms; (C) 00 levels.
  • (C) then (B), the federated-to-delegated hand-over. Three clients upload encrypted general-axis gates; the server forms the encrypted aggregate QQ and composes it with an encrypted pad key tt without decrypting or normalising: 0+10 + 1 levels, ring 2142^{14} (OpenFHE rejects 2132^{13} at this depth), 7272 ms, key recovered to 3.4×10−73.4 \times 10^{-7} rad (8.7×10−98.7 \times 10^{-9} rad at 45-bit scaling). The server uses the conjugate Qˉ=∥Q∥2Q−1\bar{Q} = \|Q\|^2 Q^{-1}, so t⊗Qˉt \otimes \bar{Q} is a positive multiple of the exact update t⊗Q−1t \otimes Q^{-1}; the key holder's normalisation after decryption removes the factor (proof in Appendix G.6). Evaluating the exact reciprocal homomorphically instead costs 88 levels.
  • LL private gates ×\times encrypted key: exactly ⌈log⁡2(L+1)⌉\lceil \log_2(L+1) \rceil levels, measured 1,2,3,41, 2, 3, 4 for L=1,2,4,8L = 1, 2, 4, 8, all at ring 2142^{14} (0.060.06–0.930.93 s).

Who holds rr. In the federated loop the server performs only (C) and never applies rotations. (A) and (B) occur only when a client delegates circuit execution to a remote processor. The "optional plaintext rotation V−1V^{-1}" in the protocol box had no role in federated optimisation and is removed from §4.2 and Figure 2.

Novelty and complexity advantage over Ma & Li (2022) (also Q3). The quaternion encoding, its degree-22 key update (their Eq. 2.14, Remark 4.1) and the quaternion/Pauli switching (their §3.2, Prop. 4.3, Lemma 5.1) are theirs, and the revised paper says so throughout. Their scheme is single-client and theoretical: no federated protocol, no aggregation geometry, no CKKS instantiation, no implementation. This paper adds the federated protocol; the costs of (B), (C) and their composition; branch/hemisphere handling and the exact aggregation bias (Prop. 2); and the first implementation.

The measured advantage over angle coordinates: from encrypted Euler angles the same key update needs 77 further levels (Chebyshev trigonometry and quaternion assembly) at every LL, at ring 2152^{15}, and 1919 s against 0.90.9 s at L=8L = 8. We also report the case that is less favourable to us: for coaxial gates angles add under encryption, so the angle chart needs 66 levels whatever LL — more than the quaternion tree up to L=31L = 31, fewer from L=64L = 64.

On practicality. The ≈1.5\approx 1.5 h per encrypted CNOT comes from the single-threaded Boolean sub-layer of the bit-level reference on the delegation channel; the federated loop never evaluates a CNOT homomorphically. In CKKS, the entire per-round homomorphic work of E7 is 1.91.9 ms of encryption per client, 0.70.7 ms of aggregation and 0.80.8 ms of decryption. The delegation cost is stated as a limitation.

AuthorsOpenReview ↗

Response to Reviewer aDWa (Part 2/2)

W2 / Q2 — threat model and confidentiality between clients. You are right: "peers see the decrypted aggregate and nothing further by protocol" is not a cryptographic guarantee under a shared key. The revised §2 and Table 14 state the cross-silo HE-FedAvg model (as in BatchCrypt, Zhang et al. 2020): an honest-but-curious server that does not collude with clients; uploads on confidential point-to-point channels; clients hold the decryption key, the server only public and evaluation keys. Under these assumptions a curious client sees only the decrypted aggregate; without them it could decrypt an intercepted upload.

To remove the shared key, we evaluated NN-out-of-NN threshold CKKS (Mouchet et al. 2021; OpenFHE multiparty): decrypting an upload with a single client's share fails; the aggregate decrypts from all shares to within 4.3×10−74.3 \times 10^{-7} rad; the cost is one extra exchange and a 263263 kB partial decryption per client and round (E9). Who can decrypt: in the shared-key variant, any client and never the server; in the threshold variant, only all clients jointly, and only the aggregate.

Q3 — which operations are homomorphic.

Exp.Backend, securityHomomorphicPlaintextHardware
E7, E7b, E7cCKKS, 128-bitserver aggregation, every roundtraining, decryption–
E8, E8b, E9CKKS, 128-bitCases A/B/C, aggregate ×\times key, product trees, threshold aggregation––
E12CKKS, 128-bitaggregation and composition of SO(3)SO(3) rotationslocal training–
E1, E5, E6bit-level LWE, structuralper-client encrypt/decrypt round triptraining; masked secure aggregation of angles–
E2, E3bit-level LWE (E2 also CKKS)aggregation (E2), round trips (E3)reference means–
E4bit-level LWEpad, key updates, 22 encrypted CNOTs–simulator
E11ibm_fez; CKKS, 128-bitpad-key update (Case A), matching the reference to 1.2×10−101.2 \times 10^{-10}double-precision reference compiled into the circuitspad, RXR_X, unpad, X/Y/ZX/Y/Z tomography
Functional check (6464 shots)bit-level LWEpad-key update–pad, RXR_X, unpad, measure

E1/E5/E6 used the harness; E7 is the primary end-to-end result (Table 1). The exact statevector simulation used in E7 matches Qiskit's EstimatorQNN parameter-shift evaluation of the same circuit to 4.4×10−164.4 \times 10^{-16} in outputs and in all 1,0041{,}004 gradients (Appendix G.4).

Q4 — security of the experimental configurations. Security claims rest on the CKKS instantiation: every context of E7–E9 and E12 runs under OpenFHE's HEStd_128_classic guard, which refuses non-compliant parameters (ring 2132^{13} for aggregation, 2142^{14} for depth-11 composition). For the bit-level reference, Appendix J now contains an executable core-SVP audit (primal uSVP and dual attack; 0.292b0.292b classical, 0.265b0.265b quantum), validated against the block sizes published for NewHope, JarJar and BCNS. The n=1n = 1 profile falls to exhaustive search; n=64n = 64 and n=256n = 256 at q=232q = 2^{32} are broken at the smallest block size the model covers; extrapolating the scheme's own noise rule, n=1024n = 1024 gives about 115115 classical bits and n=2048n = 2048 about 294294. No security is claimed for any bit-level run; those runs establish structure, exactness and cost only. The audit is a conservative estimate, not an exact security level.

Ethics flag. Thank you. An AI Use Statement appears in the main paper after the Ethics statement. In full: LLMs were used as collaborative assistants, under the direction, verification and audit of the authors, who take full responsibility for all content, mathematical claims, experimental methods, implementation choices and conclusions. Specifically: writing (assisting in drafting, restructuring and editing the manuscript); framing and mathematics (assisting in structuring the coordinate-chart framing, formulating the three-case ledger and Corollary 1, and drafting proof steps, each verified by the authors); experiment design and analysis (assisting in designing the paired-seed protocols, the branch-cut, general-gate, aggregate-then-delegate and SO(3)SO(3) experiments, the lattice-security audit and the interleaved hardware tomography, and in statistical analysis); code (assisting in developing the bit-level implementation and in writing and executing the scripts of the CKKS benchmarks, the verification and lattice audits and the hardware notebook, all audited by the authors); literature and data (assisting in retrieving and verifying references and in writing the scripts that sample seeded angles and client dispersions; no data presented as measured was generated by an LLM).

We hope these changes make every claim precise and checkable, and would be grateful if you would reconsider your assessment in their light.

Thanks for the author's response. The additional explanations and experiments help me clarify the actual contributions and positioning of this paper.

The authors have separated the cost of federated aggregation from that of delegated quantum execution. However, while the former is now demonstrated efficiently using CKKS, the latter remains substantially limited by the unimplemented CKKS track-switching mechanism and the prohibitive cost of the current bit-level reference implementation. Consequently, the paper demonstrates efficient encrypted federated parameter aggregation, but not yet an efficient end-to-end delegated execution pipeline for general quantum neural networks

The new experiments establish that circular averaging is sufficient for coaxial rotations, whereas general SU(2) transformations require a richer representation. However, the reported general-gate experiment primarily demonstrates representability rather than a practical learning advantage.

Although the author has refined the presentation and addressed some of my concerns, this also reveals the limited contributions of this paper. After explicitly acknowledging that the core mechanism of quaternion QHE originates from Ma & Li (2022), the remaining contributions mainly consist of extending the existing mechanism to federated aggregation, analyzing the associated aggregation geometry, and providing an implementation and empirical evaluation. Meanwhile, the advantages of quaternions still depend on the representation scope and specific tasks; the SU(2) structure itself does not necessarily yield better machine learning performance, which further weakens the impact of this paper from the ML perspective. While the overhead of encrypted CNOT operations does not affect the demonstrated federated aggregation pipeline, it remains a significant obstacle to practical full delegated execution.

Overall, I appreciate the authors' efforts in addressing the reviewers' concerns and acknowledge the substantial improvements in technical clarity and experimental validation. Nevertheless, my concerns regarding the novelty, broader significance, and practical scope of the contribution remain. I therefore maintain a negative recommendation, but raise my rating from Clear Reject to Weak Reject in recognition of the authors' meaningful revisions. This is my final assessment.

AuthorsOpenReview ↗

Reply to Reviewer aDWa

Thank you for engaging closely with the revision and for raising your score. For the benefit of the ACs and fellow reviewers, we briefly address the conceptual points in your final assessment:

1. Role of the SU(2)\mathrm{SU}(2) chart in ML. A homomorphic parameter chart is a cryptographic and geometric primitive, not a regulariser designed to improve plaintext accuracy over simpler ansätze. Its purpose is to encrypt a chosen architecture with zero utility loss and minimal multiplicative depth. Whenever a circuit requires general SU(2)\mathrm{SU}(2) gates (or classical SO(3)\mathrm{SO}(3) rotation layers (Cor. 1, E12)), two-component circular charts cannot represent the weights (E7c), and principal-branch angle averaging suffers catastrophic ±π\pm\pi wrap-around corruption (E7b). The ML contribution shows that non-abelian rotation parameters can be aggregated homomorphically at depth 00 and composed at depth 11 with zero utility tax (E7) and one-quarter the geometric bias of projected matrix averaging (E12).

2. Novelty beyond Ma & Li (2022). They formulated single-client quaternion QHE theoretically using an mm-bit principal-branch angle encoding. Used directly in federated learning, that encoding fails across the ±π\pm\pi branch cut (E7b). Extending the algebra to multi-client learning required results absent from prior work:

  • Aggregation geometry (Prop. 2, E12): Proving the exact cubic chordal-to-geodesic bias −124∑kwkδk3+O(δ5)-\frac{1}{24}\sum_k w_k \delta_k^3 + \mathcal{O}(\delta^5) alongside zero-depth client-side hemisphere alignment.
  • Unnormalised aggregate-to-key composition (Prop. 1, E8b): Composing an unnormalised encrypted aggregate QQ with an encrypted pad key tt via the scaled conjugate Qˉ=∥Q∥2Q−1\bar{Q} = \|Q\|^2 Q^{-1}, deferring scale removal to post-decryption key normalisation and avoiding an 88-level homomorphic reciprocal or 8–118\text{--}11-level Chebyshev trigonometry.
  • Packed CKKS protocol (E7, E9): Replacing bit-level encryption with a 128128-bit CKKS and threshold-decryption pipeline.

3. Delegated execution scope. The revised paper separates parameter aggregation (depth 00) and single-qubit rotation updates (depth 11, or ⌈log⁡2(L+1)⌉\lceil \log_2(L+1) \rceil for LL private gates) from multi-qubit track switching across encrypted CNOT\mathrm{CNOT}s, which requires a Boolean/TFHE-class backend and is stated as future work in the Limitations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the Area Chair

We thank the reviewers! Three concerns are shared: (i) attribution to Ma & Li (2022); (ii) what is encrypted, what each operation costs, and under which threat model; (iii) whether training ever ran through the protocol's homomorphic aggregation. The revision addresses them and adds new 128-bit CKKS and IBM hardware experiments (E7, E7b, E7c, E8b, E11, E12) establishing end-to-end training parity, general SU(2)SU(2) gate representability, depth-11 aggregate-then-delegate composition, and interleaved 33-basis hardware tomography.

Attribution. The quaternion encoding of SU(2)SU(2) gates and keys, its degree-22 key update and the quaternion/Pauli switching are due to Ma & Li (2022); the abstract and introduction now say so, as do Figure 1, Related Work and §4.1. The federated scaffold of our bit-level code builds on the open-source code of Dutta et al.(2024a), as the Reproducibility statement says.

Contribution. On top of Ma & Li's single-client scheme: (1) a non-interactive federated protocol in which the server aggregates all clients' SU(2)SU(2) parameters homomorphically in the quaternion chart at depth 00, and the aggregate composes with encrypted pad keys at depth 11; (2) a three-case cost ledger (Prop.1), the exact chordal-vs-Euclidean bias −(1/24)∑wδ3-(1/24)\sum w\delta^3 (Prop.2), entangler compilation, group-valued layers (Cor.1); (3) the first implementation and evaluation.

Evidence (OpenFHE CKKS, HEStd_128_classic; five paired seeds wherever a model is trained, fifteen in E7b). P = proven, M = measured.

  • E7 (M): end-to-end training, server-side homomorphic aggregation every round. IID: plaintext MSE 0.3236±0.03710.3236 \pm 0.0371 (R2=0.66R^2 = 0.66); quaternion HE-FedAvg Δ=−8.9×10−8\Delta = -8.9\times 10^{-8} (95% CI [−4.2,+2.4]×10−7[-4.2, +2.4]\times 10^{-7}); equivalent within ±0.01\pm 0.01 (TOST) under geographic and label-skew partitions and on Wine Quality.
  • E7b (M), branch cut: isolates the condition under which angle averaging fails across the ±π\pm\pi cut. On unconstrained real-valued angles sharing one lift, Euclidean averaging is unaffected (∣Δ∣≤3×10−5|\Delta| \le 3\times 10^{-5}). When angles are confined to the principal branch (−π,π](-\pi, \pi] — as in Ma & Li's mm-bit fraction-of-a-turn encoding or any atan2\mathrm{atan2} recovery — clients straddle the cut in 27–29% of rounds, every straddling parameter is corrupted by 2π/3≈2.092\pi/3 \approx 2.09 rad (7878 and 6262 times) and final MSE degrades in 2727 of 3030 runs (Wilcoxon p=0.002p = 0.002 and 6.1×10−46.1\times 10^{-4}); hemisphere-aligned quaternion HE-FedAvg suffers no corruption.
  • E7c (M), general SU(2)SU(2) gates: all four quaternion components in use (174/180174/180 trained gates); quaternion HE-FedAvg vs plaintext Δ=−5.0×10−6\Delta = -5.0\times 10^{-6}; projecting the gates onto one axis, all a two-component chart can hold, raises MSE from 0.310.31 to 0.870.87.
  • E8/E8b (P, M): encrypted aggregate ×\times encrypted key, without decrypting or normalising: 11 level, ring 2142^{14}, 7272 ms, 3.4×10−73.4\times 10^{-7} rad. LL private gates ×\times key: exactly ⌈log⁡2(L+1)⌉\lceil \log_2(L+1) \rceil levels (1,2,3,41, 2, 3, 4 for L=1,2,4,8L = 1, 2, 4, 8). Angle charts: 66 levels (coaxial) and 88–1111 (Euler), ring 2152^{15}.
  • E9 (M): one 132132 kB ciphertext per client-round; threshold CKKS: a single share cannot decrypt an upload.
  • E12 (M): classical SO(3)SO(3) federated benchmark: 1616 vs 2727 products per encrypted composition; cubic law confirmed (exponent 3.003.00; projected matrix mean 4.01×4.01\times the bias).
  • Verification (M): the exact statevector simulation matches Qiskit's EstimatorQNN parameter-shift evaluation to 4.4×10−164.4\times 10^{-16} in outputs and in all 1,0041{,}004 gradients. A core-SVP audit, validated on the published NewHope/JarJar/BCNS block sizes, shows that the bit-level profiles (n≤256n \le 256) give no security; none is claimed.

Hardware. In the submission, the encrypted hardware point (FZ=0.9918F_Z = 0.9918) came from one 6464-shot execution at θ=1.721\theta = 1.721 rad, compiled without stage barriers, and the unencrypted control from a separate, non-interleaved run at 40964096 shots. Because mismatched shot budgets and ZZ-basis-only measurement cannot isolate the physical overhead of the pad, we replaced that comparison with an interleaved experiment on ibm_fez (E11, Appendix G.14; M): 2020 angles ×\times 33 interleaved arms (unencrypted, padded, pad-only) ×\times X/Y/ZX/Y/Z state tomography at 40964096 shots on physical qubit 114114, with stage barriers and no readout mitigation. State fidelities are 0.98500.9850 (unencrypted), 0.98120.9812 (padded) and 0.99730.9973 (pad-only): carrying the quaternion pad through the rotation costs 3.8×10−33.8\times 10^{-3} of state fidelity (95% CI [0.8,6.7]×10−3[0.8, 6.7]\times 10^{-3}, p=0.014p = 0.014), well inside the pre-specified ±0.01\pm 0.01 margin (TOST p=1.2×10−4p = 1.2\times 10^{-4}). The CKKS Case A key update (11 level, ring 2142^{14}) reproduces the unpad unitaries executed on ibm_fez to 1.5×10−101.5\times 10^{-10} in operator norm.

Title. We propose "Homomorphic Federated Learning of Quantum Neural Networks in Quaternion Chart".

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.