acceptodds
Under review as a conference paper at ICLR 2027

Trust-Aware Prototype-Guided Quantum Representation Learning for Backdoor-Resistant Medical Image Classification

Abstract

Training-time backdoors turn small image artifacts into prediction shortcuts, and compact medical classifiers are especially exposed because a low-dimensional representation leaves little room to absorb a spurious direction. We present TAPQ-Net, a hybrid quantum–classical defense that trains directly on poisoned data, without a trusted clean subset, sample removal, or post-hoc purification. A frozen encoder and a fixed PCA projection map each image to eight coordinates. A label-free expectation–maximization (EM) mixture assigns every sample a soft prototype and a trust score; an 8-qubit tree-structured quantum encoder (TQE) with seven trainable rotations is aligned to that prototype, and the trust score bounds the sample's cross-entropy weight. The mechanism that separates TAPQ-Net from prior prototype- and clustering-based defenses is a reverse channel: the EM state may evolve only when the silhouette geometry of the quantum embedding stays separated, balanced, and non-collapsed. Because EM never reads labels, relabeling reaches the prototypes only through the trigger's latent offset, and alignment bounds the trigger-direction component of the embedding. On OASIS and six MedMNISTv2 datasets under random-class and targeted patch poisoning at 1–10%, TAPQ-Net attains the lowest attack success rate in all 42 conditions (0.13–0.24, against at least 0.34 for every compared method) and the highest clean accuracy in 40, and paired tests over the 42 conditions are significant for all four metrics. A topology- and parameter-matched classical tree trails the TQE by 3.4 accuracy points, poisoned samples receive lower trust without being removed, and the defense carries over to clean-label and warping-based triggers.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Backdoor defense, medical image classification, training-data poisoning, 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_HzkWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes TAPQ-Net for medical image classification under backdoor attacks. The model builds prototypes with EM clustering and does not use labels for this step. These prototypes give each training sample a trust weight. A tree-structured quantum encoder produces the image features. The key idea is a gate that checks the geometry of the representations and decides whether to update the prototypes. The experiments use seven datasets. However, the paper does not clearly show what the quantum encoder adds. It also does not show that the gate helps beyond simply running more EM iterations.

Critical strengths
  1. The design fits together well. The label-free prototypes add a signal that the classification loss does not provide. Soft weighting keeps every sample in training instead of dropping suspected ones. The paper also separates how confident a cluster assignment is from whether a sample is clean. This is a useful distinction.
  2. The evaluation of patch attacks is broad. It covers seven datasets and several poisoning rates. The authors also analyze trust scores, compression, and triggers.
  3. The paper includes relevant controls and is open about its limits. The comparison with a matched classical tree is useful. The authors state that all quantum results come from simulation and that some controls are still missing.
Critical weaknesses
  1. The value of the quantum module is unclear. Every defense component could run with a classical encoder. The theory gives no robustness mechanism that is specific to quantum computing. Table 3 shows that the chosen TQE beats some other designs. The matched classical tree, however, appears in only one setting, and the paper does not fully describe how it is built. The authors should compare against stronger classical encoders in more settings.
  2. The novelty and effect of the gate are not well supported. Clustering, prototype supervision, sample reweighting, and dynamic prototype updates all exist in prior work. The main new element is that geometry decides when to refresh. In Table 2, though, the effect of gating cannot be separated from the effect of extra EM iterations. Three controls would separate them: always refreshing, starting from fully converged EM, and matching the number of updates.
  3. The gate may freeze permanently. The balance and activity checks rely on EM assignments. These assignments do not change when a refresh is rejected. So once a check fails, it seems that no later update can pass. The authors should say whether this happens and how often.
  4. The random-class ASR results need explanation. Suppose each test target is drawn uniformly from the wrong classes. Then the expected ASR equals the triggered error rate divided by C − 1. Several values in Table 12 are well above this level. The authors should explain how targets are chosen and how the metric is computed.
  5. The theory does not show that predictions are robust. Proposition 4.1 simplifies the EM weights, although in practice they depend on the data. Proposition 4.3 gives only a general bound on alignment. Neither result shows that predictions stay stable or that ASR stays low. That would need extra assumptions about trigger directions, classifier norms, or margins.
  6. Some details needed for reproduction and evaluation are missing. The paper does not give key hyperparameter values, the exact classical tree function, or how much clean data each baseline uses. Model selection uses a clean validation set. The statistical tests run over related conditions, so their results should be read with care.
Questions for authors
  1. Which quantum property explains the gains? Do the gains hold against fully described, strong classical encoders in several settings?
  2. How does the gate compare with always refreshing, with converged EM at the start, and with schedules that match the number of updates? How many updates does the gate accept?
  3. Can refresh restart after a balance or activity check fails? If not, how often does the gate freeze permanently?
  4. How are the random-class targets generated? How do the reported ASR values follow from the stated definition?
  5. Can the authors give full configurations, the clean data budget of each baseline, and the ASR without any defense to confirm that the attacks work?
  6. What evidence links alignment to stable outputs? Useful evidence would include measured residuals, classifier norms, or how the logits change between clean and triggered inputs.
4 comments

Thank you for the questions. Percentages below average seeds 11, 22 and 33. F1 denotes macro-F1.

Q1. Quantum mechanism and classical encoders. The TQE provides a fixed hierarchy of bounded trigonometric interactions controlled by seven learned angles. We test this feature-map design against fully specified classical alternatives on three datasets under A2A 10%:

EncoderPathMNIST F1/ASRDermaMNIST F1/ASROASIS F1/ASR
TAPQ-Net87.13 / 0.6157.78 / 0.5739.35 / 0.96
Scalar tree87.32 / 0.4158.93 / 0.5437.90 / 0.54
MLP87.14 / 0.4659.28 / 0.2837.57 / 0.83
Random Fourier86.77 / 0.5458.65 / 0.3538.40 / 0.75
Exact recurrence86.31 / 0.5759.79 / 0.4840.26 / 0.98

The exact classical recurrence matches circuit expectations to 8.88×10−168.88\times10^{-16} at identical parameters; the table compares independently trained endpoints. Classical alternatives are competitive. We specify all nine encoders' functions, dimensions, operation/parameter counts and refresh counts.

Revision: Sec. 5.4, pp. 8–9; Table 5, p. 9; Appendix A.1, p. 14; Appendix B.4, pp. 17–18; Table 19, p. 23.

Q2. Gate attribution and EM controls. All 126 primary full runs accept every eligible update (7/7). Their refresh epochs match the fixed interval, so the primary-grid gains cannot be attributed to vetoes. PathMNIST A2A 10% gives:

ConfigurationMacro-F1ASRLater updates
TAPQ-Net87.130.617
Frozen EM86.610.400
Always refresh86.210.3115
Fixed interval85.910.457
Matched count86.580.187
Converged initialization86.660.350
Shuffled monitor86.700.487
Silhouette only86.460.407
Balance/activity only85.610.417
Without balance86.820.467
Without activity86.740.437

These compare complete configurations with shared initialization records and different circuit-evaluation pathways. Converged initialization adds 142/148/139 initial EM passes (14.2/14.8/13.9 s). The gate implements the stated geometric acceptance rule.

Revision: Sec. 5.3, pp. 7–8; Table 4, p. 9; Appendix D.2, pp. 22–23; Tables 17–19, p. 23.

Q3. Restart and permanent freezing. There is no restart after an occupancy failure, and zero premature permanent freezes occur in the 126 primary full-model runs (882/882 checks accepted). Retaining the mixture preserves hard assignments and their balance/activity statistics, so such a rejection is absorbing. Silhouette-only rejection can recover as the encoder changes. Alternative encoders' 0–7 accepted updates are reported without inferring which criterion failed.

Revision: Sec. 3.5, p. 4; Appendix D.2, pp. 22–23; Table 19, p. 23.

Q4. Targets and ASR. Training and triggered testing share cyclic g(y)=(y+1) mod Cg(y)=(y+1)\bmod C for A2A or target 0 for A2O. Targets are deterministic; the independently uniform wrong-target identity does not apply. A2A uses all test images; A2O excludes true class 0.

Revision: Sec. 3.1, p. 3; Appendix B.6, p. 19; Table 9, p. 18.

Q5. Configurations, clean data and attack validity. TAPQ-Net uses no trusted clean training subset and evaluates epoch 15 without validation selection. The nominal 5% clean subsets for NP/SAU/TSBD/SCAn and every method's stages are specified; configurations were fixed, with zero post-hoc tuning runs. A2O 10% attack ASRs averaged over the three focal datasets are:

AttackNo defenseTAPQ-Net
Default patch99.840.50
SIG73.450.37
WaNet68.770.38
Frozen-feature latent separability99.860.66
Joint classifier-and-gate99.900.77

Revision: Tables 8–10, pp. 17–19; Appendix B.5, pp. 18–19; Appendix B.7, p. 19; Appendices E.1–E.2, p. 24; Tables 6 and 20, pp. 10 and 25.

Q6. Alignment and output stability. At the auxiliary epoch-14 encoder/head checkpoint, 99.05%–99.43% of paired predictions are unchanged, using frozen epoch-15 mixture targets:

DatasetResidual clean/triggeredMean max logit changeUnchanged predictions
PathMNIST0.042 / 0.0480.11899.43%
DermaMNIST0.068 / 0.0750.14699.42%
OASIS0.054 / 0.0610.09499.05%

Table 7 gives exact changed counts per seed. Both inputs use the same epoch-14 encoder/head; the primary endpoint remains epoch 15. Proposition A.1 (p. 15) gives prediction preservation when every clean logit margin exceeds ∥wy−wk∥2Dz\|w_y-w_k\|_2D_z; DzD_z uses alignment residuals, target-direction and norm changes. The prototype bound retains fitted data-dependent weights. Statistics average related conditions within seeds (Appendix H, pp. 26–27).

Revision: Sec. 4, p. 5; Appendix A.2, pp. 14–15; Appendices A.3–A.4 and Table 7, p. 15.

Thank you for the added experiments and explanations. Reviewer oE5X and reviwer 12t3 also questioned the gate’s contribution, and I share this concern. Using a fixed training epoch addresses the clean-validation concern. The attack targets are now clearer, and the new adaptive attack results are useful.

My main concern still remain the gate’s contribution and the need for the quantum circuit. The gate is meant to stop potentially harmful prototype update, but it never reject an update in any primary run. How do these experiments show that it help? Since the update schedule is identical to fixed-interval refresh, why do the results differ? The full model and control use different circuit-evaluation pathways, so the current comparison do not answer this question. Please keep the implementation the same and compare results with the gate on and off. The classical model is also competitive in the new experiments. It is still unclear what extra benefit the quantum circuit provides.

Thank you for recognizing the fixed epoch-15 clarification, the target definitions and the added adaptive-attack results. We address the gate and encoder questions separately below.

Part 1/2: Gate contribution, refresh-schedule results and the requested gate-on/off comparison

We present the gate as a conditional prototype-refresh mechanism and distinguish its observed decisions from the performance of the complete training configuration. The primary runs accept every eligible check, so their gains cannot be attributed to refresh vetoes.

A. What the experiments establish about the gate

  1. Observed behavior in the primary runs. All 126 full-model runs accept all 882 eligible checks, giving seven refreshes per run at epochs 2, 4, 6, 8, 10, 12 and 14. The observed premature-stop count is 0/126. These runs therefore follow the fixed-interval refresh schedule (§5.3, p. 8; Appendix D.2, p. 23).

  2. Conditional behavior across encoder configurations. Under the acceptance rule in Eq. (8), p. 4, Table 19, p. 23, reports mean accepted-update counts of 7.0 on each focal dataset for TQE and exact recurrence, and 0.0 for scalar tree, MLP and random Fourier features. Other means include 6.7 for no-CNOT and flat-VQC configurations on OASIS, 6.3 for the four-qubit tree on DermaMNIST, and 6.7/6.0 for the deeper tree on PathMNIST/OASIS. These are acceptance trajectories; the counts do not identify the failed criterion or establish that a rejected update was harmful.

  3. Behavior under the joint adaptive attack. The classifier-and-gate attack retains seven accepted refreshes per seed and dataset. TAPQ-Net ASRs are 0.48/0.64/1.18% on PathMNIST/DermaMNIST/OASIS, versus 99.97/99.74/100.00% undefended. The reported three-dataset means are 0.77% and 99.90%. These results measure the complete defense while the gate continues to pass (Table 6, p. 10; Appendix E.2, Eq. (23), p. 24).

The gate tests embedding silhouette and mixture occupancy before another EM pass. As specified in §3.5, p. 4, an occupancy failure is absorbing without a restart, whereas a silhouette rejection can recover as the encoder changes. We report these operational properties without treating acceptance as proof of semantic correctness.

B. Why the full-model and fixed-interval scores differ

Table 4, p. 9, documents uncached statevector evaluation for TAPQ-Net and cached circuit evaluation for the refresh controls. Appendix B.7, p. 19, records shared PathMNIST features/PCA, initial encoder/head and mixture states, poison indices and minibatch ordering. The comparison therefore matches these initial records but changes the evaluation pathway. The reported results do not isolate the numerical cause of the endpoint gap.

C. Numerical comparison and status of the requested control

The following are three-seed mean Macro-F1 / ASR (%) under A2A 10% poisoning. Seeds are 11, 22 and 33. U is the number of PathMNIST post-initialization updates. The first row uses uncached evaluation; the listed controls use cached evaluation.

ConfigurationPathMNISTDermaMNISTOASISPathMNIST U
TAPQ-Net87.13 / 0.6157.78 / 0.5739.35 / 0.967
Fixed interval85.91 / 0.4558.39 / 0.5940.27 / 0.907
Always refresh86.21 / 0.3158.54 / 0.4539.71 / 0.4615
Matched pass count86.58 / 0.1860.96 / 0.5438.50 / 0.617
Alignment, frozen EM86.61 / 0.4058.52 / 0.4040.00 / 0.570
Converged initialization86.66 / 0.3557.45 / 0.4939.81 / 0.280
Shuffled monitor86.70 / 0.4857.53 / 0.3140.21 / 0.497
Silhouette only86.46 / 0.4056.02 / 0.4139.29 / 0.677
Balance/activity only85.61 / 0.4160.14 / 0.4439.00 / 0.567

Locations: Table 4, p. 9; Appendix D.2, Table 17, p. 23.

Table 18, p. 23, additionally reports PathMNIST no-balance and no-activity scores of 86.82/0.46 and 86.74/0.43, respectively, with seven accepted updates in every seed.

Part 2/2: Motivation and measured contribution of the tree encoder

We use the TQE to specify a compact hierarchy of interactions among the eight compressed image-feature coordinates. Seven learned angles couple local coordinates into progressively larger branches, with eight bounded readouts available for prototype alignment and classification. Four readouts preserve individual-coordinate cosines; two combine pairs, one combines four inputs, and the root combines all eight. This defines the feature family using a small number of learned coefficients (§3.3, pp. 3–4; Appendix A.1, p. 14).

Including the biased linear head, the encoder has 7+9C7+9C gradient parameters: 88/70/43 on PathMNIST/DermaMNIST/OASIS. The scalar tree has the same counts. The compact MLP has four hidden outputs and aligns to the first four prototype coordinates; it is an alternative representation, not an exact equal-capacity substitution (Appendix B.4, pp. 17–18).

The comparisons show task-dependent utility and ASR trade-offs. Scores below are Macro-F1 / ASR (%) under A2A 10% poisoning, averaged across seeds 11, 22 and 33. Parameter counts exclude the fitted mixture and frozen backbone; U gives mean post-initialization refreshes. Triplets use PathMNIST/DermaMNIST/OASIS order.

EncoderPathMNISTDermaMNISTOASISGradient parametersU
TAPQ-Net87.13 / 0.6157.78 / 0.5739.35 / 0.9688/70/437.0/7.0/7.0
Exact recurrence86.31 / 0.5759.79 / 0.4840.26 / 0.9888/70/437.0/7.0/7.0
Scalar tree87.32 / 0.4158.93 / 0.5437.90 / 0.5488/70/430.0/0.0/0.0
MLP87.14 / 0.4659.28 / 0.2837.57 / 0.8381/71/560.0/0.0/0.0
Random Fourier86.77 / 0.5458.65 / 0.3538.40 / 0.7581/63/360.0/0.0/0.0
No CNOT85.95 / 0.4358.49 / 0.2839.66 / 0.8488/70/437.0/7.0/6.7
Flat VQC87.11 / 0.4859.93 / 0.3238.54 / 0.4588/70/437.0/7.0/6.7
Four-qubit tree86.56 / 0.3459.85 / 0.4639.57 / 0.5148/38/237.0/6.3/7.0
Deeper tree86.17 / 0.5656.88 / 0.4338.98 / 0.2695/77/506.7/7.0/6.0

Locations: Table 5, p. 9, for scores; Appendix D.3, Table 19, p. 23, for parameter counts and mean refresh counts.

On OASIS, TQE's Macro-F1 exceeds scalar tree, MLP and Fourier features by 1.45/1.78/0.95 points, with higher ASR. Classical alternatives are competitive on PathMNIST and stronger on DermaMNIST. The deeper tree has lower Macro-F1 on all three datasets, accompanied by lower ASR. These are results for the trained configurations and their resulting refresh trajectories; acceptance counts do not establish superior semantic geometry.

Quantum execution is not required for this feature map. Equation (12), p. 5, gives its exact seven-merge classical recurrence. Appendix A.1, p. 14, verifies it against independent statevector evaluation over 100 random forward cases and 70 finite-difference gradient columns, with maximum discrepancies 8.88×10−168.88\times10^{-16} and 3.33×10−103.33\times10^{-10}, respectively. It is the same function at identical inputs and parameters; independently trained endpoint differences do not demonstrate a quantum-execution benefit.

Appendix B.4, p. 17, counts 35 multiplications and 14 additions/subtractions for the merges, excluding trigonometric evaluation and the common head. This gives a classical evaluation route without statevector simulation; it is not a zero-cost or measured deployment-speed result.

The noise analysis has a specified scope. Proposition A.3, p. 16, shows that a common positive final-readout contraction preserves cosine alignment and silhouette at fixed parameters, targets and assignments with inactive norm floors. A classifier bias can still change predictions. Appendix F, p. 25, and Figure 6, p. 27, separately evaluate simulated 256/1,024/8,192-shot and gate-level channel configurations. Those measurements do not constitute hardware validation or general noise immunity.

The evidence therefore supports evaluating this seven-angle feature family within prototype-guided training. It does not establish a requirement for quantum execution or consistent superiority over the classical alternatives. Thank you for helping us make that contribution and the gate-attribution boundary precise.

Reviewer_oE5XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

TAPQ-Net uses a frozen ResNet-18, a PCA that has been trained, a label-free Gaussian mixture, and a small trainable tree quantum encoder in order to learn from backdoor-poisoned medical images. The mixture responsibilities generate soft prototype targets and assignment-concentration scores. These targets supervise the embedding by means of cosine alignment, while the concentration scores are used to weight the cross-entropy. The decisive component is a feedback rule according to which the silhouette score of the learned embedding, together with the mixture balance and activity, decides whether or not the mixture proceeds to the next EM step.

The paper provides conditional geometric bounds and includes experiments involving seven medical datasets, two attack-label mappings, and three poisoning rates, together with additional studies on the encoder, the trust score, compression, and the trigger.

Critical strengths
  1. The method is coherent and makes a reasonable specific contribution. Soft responsibilities are applied consistently when setting up the targets and the weight for supervision, with all the training samples being kept in. The authors accurately define trust as being about assignment concentration rather than as a calibrated probability of cleanliness; authors also place their various ideas in contrast to previous approaches based on clustering, prototyping, and quantum methods.
  2. Manuscript has provided evidence along with some useful controls. The evaluation based on seven datasets involves three runs, several metrics, a comparison with a classical tree, trust diagnostics, and a PCA sweep. Authors have provided a clear discussion of the clean validation, the simulation, and the missing gate controls.
Critical weaknesses

W1. The central ablation does not isolate the distinctive gate. As shown in Table 2, going from (e) to (f) increases ACC from 80.6 to 89.5 and lowers the reported ASR from 0.29 to 0.16. Yet (e) freezes the mixture after the first EM pass, while (f) both carries out EM refinement and decides when to do so based on the embedding geometry. This shows an advantage of the complete refresh procedure over the frozen initialization; it does not prove that the conditional rule performs better than ordinary refinement. This limitation is acknowledged in the main text (pages 8-9) and in Apps. D.6 and G.2. Since the gate is the main distinguishing feature, it is important to carry out a focused comparison with unconditional refresh and with a comparable-update-count fixed schedule. A converged-initialization control would also help to determine whether staged refinement is necessary.

W2. The explanation of label isolation also needs precise qualification. When the features are fixed, initialization is given, and the number of accepted EM passes is fixed, the mixture state is independent of the labels. In Algorithm 2, the labels influence the encoder via the cross-entropy term, and the encoder in turn affects the silhouette test and thus the schedule for accepted updates. It refers that at a given epoch, re-labelling could choose a different prefix of the same label-free EM sequence. This does represent a possible dependency in the combined algorithm, even though it did not occur in the experiments described. Although the conditional inequalities remain useful, they alone do not prove that the trained classifier is robust.

W3. Reproducibility is limited by the lack of code and unclear implementation details. Also, the way the random-class ASR is defined needs to be made clearer in the same protocol issue.

Questions for authors
  1. Regarding gate attribution, on the current PathMNIST setup, how does the proposed gate compare with unconditional periodic refresh, a pre-specified schedule that uses a similar number of EM passes, and a mixture that has been converged before representation training. Provide the number of updates and the associated timing.

  2. For the attack mapping and evaluation, how are the training poison labels and the random-class test targets produced, and what mapping do they have in common?

  3. Regarding reproducibility and comparison, please provide the nine numerical settings, the baseline clean-data and tuning budgets, the exact number of node operations for the classical tree and the MLP, and the dataset release and split manifests. Reproducibility is limited by the lack of code and unclear implementation details.

1 comment

Thank you for the focused questions. We have added the controls and specified the implementation. Percentages average seeds 11, 22 and 33. F1 denotes macro-F1.

Q1. Updates, initialization and timing. The primary gate accepts all seven scheduled refreshes, matching the fixed interval. PathMNIST A2A 10% results are:

ScheduleF1/ASRLater updatess/epochWall s
TAPQ-Net87.13 / 0.61774.861225.8
Frozen EM86.61 / 0.40010.89185.9
Always refresh86.21 / 0.311510.49182.2
Fixed interval85.91 / 0.45710.80186.0
Matched count86.58 / 0.18710.68367.8
Converged initialization86.66 / 0.3509.70169.3

Converged initialization takes 142/148/139 EM passes and 14.2/14.8/13.9 s before training (tolerance 10−610^{-6}, maximum 500 passes). Fixed interval updates after epochs 2,4,...,14; count matching samples seven epochs from 1–15. The comparisons reuse feature/PCA projections, initial encoder/head and EM states, poison indices and minibatch order, except the deliberate additional initialization work.

Profiling uses one A100-SXM4-40GB. The full-model row uses uncached statevector evaluation; the control runs use cached circuit evaluation. Wall timers include loading, initialization, training and test inference; epoch timers include training and scheduled EM. The timing comparison therefore includes this execution-path difference. Across 126 primary full runs, all 882 eligible checks are accepted, as detailed in the refresh analysis.

Revision: Sec. 5.3, pp. 7–8; Table 4, p. 9; Appendix D.2, pp. 22–23; Appendix I, p. 28.

W2. Conditional label isolation. EM is label-free at fixed features, initialization and accepted-pass count; the full training trajectory can depend on labels. Cross-entropy changes the encoder, which changes silhouette and can select a different prefix or timing of the same EM sequence. Our prototype bound retains the fitted, data-dependent weights and makes no independence assumption for them.

Revision: Sec. 3.5, p. 4; Sec. 4, p. 5; Appendix A.2, pp. 14–15.

Q2. Training and test targets. Both use the same deterministic mapping: cyclic g(y)=(y+1) mod Cg(y)=(y+1)\bmod C for A2A and target 0 for A2O. The independently uniform wrong-target identity therefore does not apply. Poison sampling uses seed 10,000+s10{,}000+s; clean-label SIG preserves target-class training labels. A2A ASR evaluates all test images; A2O excludes true class 0. Its eligible counts are 9,000 CIFAR-10; 3,177 D1; 5,842 D2; 16,742 D3; 7,388 D4; 750 D5; 1,939 D6; 2,806 OASIS. Clean ACC uses all unmodified test images. Seeds receive equal weight.

Revision: Sec. 3.1, p. 3; Appendix B.6, p. 19; Table 9, p. 18.

Q3. Reproducibility and budgets. We provide the nine numerical settings, method-specific budgets and reproducibility records in the revised manuscript and anonymous repository:

SettingValue
Responsibility temperature τ\tau1.0
EMA factor α\alpha0.5
Minimum support mmin⁡m_{\min}10
Alignment weight λalign\lambda_{\mathrm{align}}0.5
Confidence floor rmin⁡r_{\min}0.2
Refresh interval Δref\Delta_{\mathrm{ref}}2 epochs
Silhouette tolerance εsil\varepsilon_{\mathrm{sil}}0.02
Maximum component share ρmax⁡\rho_{\max}0.5
Minimum active components Kmin⁡K_{\min}⌈K/2⌉\lceil K/2\rceil

TAPQ-Net uses Adam 10−310^{-3}, batch 128, 15 epochs and no clean training subset or validation selection. NP/SAU/TSBD use 15 base + 10 repair epochs (plus TSBD's ascent pass). They and SCAn receive 4,499/500/1,929/2,500 trusted images on D2/D6/OASIS/CIFAR-10; the last is 5% before, or 5.56% after, the validation holdout. Table 10 gives every method's stages. The configurations were fixed, with zero post-hoc tuning runs.

The scalar tree uses 8 input cosines, 7 multiplications, 7 additions and 7 tanh evaluations, plus 8C8C head MACs; its 7+9C7+9C parameters are 88/70/43 on D2/D6/OASIS. The 8→4→C8\to4\to C MLP uses 32+4C32+4C MACs, 4 tanh evaluations and 36+5C=81/71/5636+5C=81/71/56 parameters.

OASIS uses participant-stratified 70/15/15 splits (seed 0), with zero pairwise patient overlap. MedMNISTv2 uses the named tasks' loader partitions. The anonymous repository includes the software environment (Python 3.10, PyTorch 2.1.2/CUDA 12.1 and PennyLane 0.35.1), dataset and participant-split manifests, trusted-subset membership, tuning records and final epoch-15 checkpoints.

Revision: Appendices B.2–B.4, pp. 17–18; Appendix B.5, pp. 18–19; Appendix B.7, p. 19; Tables 8–10, pp. 17–19.

Reviewer_12t3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TAPQ-Net, a hybrid quantum–classical framework for defending medical image classifiers against training-time backdoor attacks without requiring a trusted clean training subset, removing suspicious samples, or performing post-hoc model purification. The method combines a frozen ResNet-18 feature extractor and an eight-dimensional PCA projection with a label-free Gaussian mixture model (GMM) that produces soft prototypes and sample-specific trust scores. An eight-qubit tree-structured quantum encoder (TQE) is trained using prototype–quantum alignment and trust-weighted cross-entropy. The central innovation is a feedback-controlled prototype refresh mechanism that allows the GMM state to evolve only when the learned quantum embedding satisfies predefined geometric criteria, including silhouette quality, component balance, and the number of active components. The method is evaluated on OASIS and six MedMNISTv2 datasets under random-class and targeted patch poisoning at rates of 1%, 5%, and 10%. Across 42 dataset–attack–poisoning-rate conditions, the authors report the lowest attack success rate (ASR) in every condition and the highest clean accuracy in 40 conditions, comparing TAPQ-Net with 11 baseline methods or reference approaches. Additional experiments investigate component ablations, alternative encoders, trust-score behavior, different trigger configurations, clean-label attacks, and warping-based attacks. The paper also provides theoretical bounds on prototype deviation and the trigger-direction component of the learned embedding under explicit assumptions. Although the empirical results are promising, the contribution of the feedback gate is not fully isolated from the effect of additional EM refinement, and the evaluation relies on a clean validation split despite the stated goal of avoiding trusted clean data.

Critical strengths
  1. A coherent and potentially useful defense mechanism. The paper addresses a relevant problem: preventing backdoor triggers from dominating compact representations in medical image classification. Combining label-free soft prototypes, trust-weighted supervision, prototype–quantum alignment, and representation-dependent prototype updates provides a coherent approach to reducing the influence of poisoned labels without explicitly filtering training samples. The feedback-controlled interaction between the learned representation and its supervisory prototypes is an interesting design choice that distinguishes the method from approaches relying on fixed prototypes or unconditional clustering updates.
  2. Broad empirical evaluation with multiple complementary analyses. The evaluation covers seven datasets, two poisoning mappings, three poisoning rates, and multiple baseline families. The reported reduction in ASR is substantial, while clean classification performance is generally maintained or improved. The authors also include ablation studies, alternative frozen backbones, trigger variations, trust-score diagnostics, and a topology- and parameter-matched classical tree comparison. These experiments provide useful evidence for the potential effectiveness of the overall framework. The paper also explicitly acknowledges several limitations, including the absence of adaptive-trigger evaluations and execution on actual quantum hardware.
Critical weaknesses
  1. The central contribution of feedback-controlled prototype refresh is not sufficiently isolated. The main novelty is the mechanism that determines whether the prototype state should be updated according to the geometry of the quantum embedding. However, the comparison between rows (e) and (f) in Table 2 changes both the presence of the feedback gate and the number of EM refinement passes. Consequently, the substantial improvement reported for the full model cannot be attributed uniquely to the proposed gate; it may partly result from performing additional prototype refinement. The authors acknowledge this limitation in Appendix D.6 and specify the missing controls in Appendix G.2. Because the feedback mechanism is central to the claimed contribution, this is more than a minor ablation issue. Without comparisons against unconditional refresh and other controls matched for the number of EM updates, the evidence does not establish that geometry-dependent feedback is responsible for the observed gains.
  2. The claimed robustness setting and the evaluated threat model are not fully aligned. Although the paper emphasizes that TAPQ-Net does not require a trusted clean subset, the implementation uses a clean validation split for epoch selection (Appendix G.5). This is a meaningful assumption for a defense intended for potentially compromised training pipelines and should be distinguished clearly from the absence of a clean training subset. Furthermore, the clean-label SIG and WaNet experiments are restricted to PathMNIST and compared only with NQNN. The paper also identifies adaptive triggers that deliberately form a compact, well-separated latent component as an important unresolved failure mode. Therefore, the current results support effectiveness under the tested attacks, but provide more limited evidence for robustness against adaptive backdoor attacks or under settings where clean validation data are unavailable.
Questions for authors
  1. Can the authors isolate the effect of the feedback gate from that of continued EM refinement? In Table 2, the comparison between the fixed-prototype alignment variant and the full model changes both the prototype-refresh mechanism and the number of EM passes. Could the authors provide an always-refresh or fixed-interval refresh baseline, with the number of EM updates matched to that of TAPQ-Net? A shuffled-gate control or separate ablations of the silhouette, balance, and active-component criteria would also help establish which aspects of the feedback mechanism are necessary. How much of the reported improvement remains attributable specifically to geometry-dependent gating?
  2. What clean-data assumptions are required by the proposed method? The paper states that TAPQ-Net does not require a trusted clean subset, but Appendix G.5 indicates that a clean validation split is used for epoch selection. Does the method require trusted, labeled validation data in addition to the poisoned training set? If so, could the authors clarify and narrow the claim accordingly? Evidence on performance when the validation split is contaminated, or when clean validation data are unavailable, would help establish the practical scope of the defense.
  3. How does the method perform against adaptive triggers designed to evade the geometric gate? The theoretical analysis identifies the possibility that a trigger may form a compact, well-separated latent component, potentially violating the assumptions underlying the prototype-deviation bound. Has TAPQ-Net been evaluated against an adaptive attack designed to preserve the embedding's silhouette score and component balance while maximizing ASR? Such an evaluation would materially strengthen the evidence for the proposed defense's robustness beyond the tested trigger configurations.
3 comments

Thank you for the questions. We have added the requested controls and clarified the protocol. Values below are percentages averaged over seeds 11, 22 and 33. F1 denotes macro-F1.

Q1. Gate versus continued EM. All 126 primary full-model runs accept all seven eligible refreshes (882/882 checks). Their update epochs therefore match the fixed interval; gains on this grid cannot be attributed to refresh vetoes. PathMNIST results under cyclic A2A patch poisoning at 10% are:

ConfigurationMacro-F1ASRLater refreshes
TAPQ-Net87.130.617
Frozen EM86.610.400
Always refresh86.210.3115
Fixed interval85.910.457
Matched count86.580.187
Converged initialization86.660.350
Shuffled monitor86.700.487
Silhouette only86.460.407
Balance/activity only85.610.417
Without balance86.820.467
Without activity86.740.437

TAPQ-Net has the highest listed Macro-F1; several controls have lower ASR. These are complete-configuration outcomes. The comparisons reuse initial feature/PCA, encoder/head and EM states, poison indices and minibatch order, but the recorded full/control runs use different circuit-evaluation pathways. Converged initialization deliberately performs 142/148/139 initial EM passes (14.2/14.8/13.9 s) before training. The gate is designed to reject geometrically inadmissible updates. Table 19 reports the resulting acceptance counts for all alternative encoders (0–7 refreshes).

Revision: Sec. 5.3, pp. 7–8; Table 4, p. 9; Appendix D.2, pp. 22–23; Tables 17–19, p. 23.

Q2. Clean-data assumptions. Our primary protocol requires neither trusted clean training samples nor clean-validation selection: it evaluates the fixed epoch-15 checkpoint. Separate checkpoint-policy comparisons under A2A 10% give:

DatasetFinal epoch F1/ASRClean-validation selectionPoisoned-validation selection
PathMNIST87.13 / 0.6186.09 / 0.5685.29 / 0.35
DermaMNIST57.78 / 0.5757.85 / 0.7057.37 / 0.37
OASIS39.35 / 0.9640.41 / 0.2838.56 / 0.55

The policies select from the same training runs. NP/SAU/TSBD/SCAn receive a nominal 5% trusted clean training subset: 4,499 PathMNIST, 500 DermaMNIST, 1,929 OASIS and 2,500 CIFAR-10 images. For CIFAR-10 this is 5% of the original 50,000-image pool, or 5.56% after the validation holdout. Membership records are archived.

Revision: Sec. 3.1, p. 3; Sec. 5.6, pp. 9–10; Appendix B.5, pp. 18–19; Table 22, p. 27.

Q3. Gate-aware adaptive attacks. The joint white-box attack reaches mean targeted ASR of 0.77% with TAPQ-Net versus 99.90% undefended, while preserving all seven hard-gate passes in every seed. Under target-0 A2O poisoning at 10%:

DatasetFrozen-feature: no defense / TAPQ ASRJoint: no-defense ASRJoint: TAPQ F1 / ASR
PathMNIST99.96 / 0.3699.9786.08 / 0.48
DermaMNIST99.62 / 0.5699.7458.52 / 0.64
OASIS100.00 / 1.04100.0038.11 / 1.18

The joint objective (Eq. 23, p. 24) combines targeted cross-entropy with soft silhouette, balance and activity penalties, accessing the frozen backbone, TQE, mixture and classifier. It uses a 3×33\times3 bottom-right mask, ∥δ∥∞≤16/255\|\delta\|_\infty\le16/255, 50 projected steps of size 2/2552/255, three restarts, and 10% poisoning (8,999/1,001/3,858 images on PathMNIST/DermaMNIST/OASIS), without covers. The frozen-feature attack is a separate 150-step Adam optimization (lr 0.1, batch 64, up to 256 non-target images), with mean defended ASR 0.66%. The broader panel includes SIG and WaNet on all three datasets against six methods.

Revision: Sec. 5.5, p. 9; Table 6, p. 10; Appendices E.1–E.2, p. 24; Table 20, p. 25.

We have incorporated these results and explanations into the revised manuscript.

Thank you for the detailed response. I appreciate the authors' efforts to address the review comments. These revisions clarify some of my concerns and improve the completeness of the paper. However, considering the paper's contributions, experimental quality, and remaining limitations, I decide to maintain my original rating.

Thank you for re-evaluating the manuscript and acknowledging that the revisions clarified some concerns. We summarize the contribution and experimental additions below.

1. Contribution and primary evidence

TAPQ-Net studies prototype-guided training of a compact representation under backdoor poisoning. Frozen image features and training-fitted PCA feed a label-free GMM that supplies soft targets and confidence weights. The encoder uses cosine alignment and weighted classification; representation geometry controls later mixture refinement. Every training example is retained. The primary endpoint is epoch 15, requiring neither a trusted clean training subset nor clean-validation checkpoint selection.

An exact classical recurrence and conditional prototype/alignment analysis characterize the seven-angle tree design.

Locations: §1, pp. 1–2; §3–4, pp. 3–5; Appendix A, pp. 14–16.

Selected condition-balanced results from Table 1 are below. Values are mean ± sample SD over seeds 11, 22 and 33; ACC, Macro-F1 and ASR cover 42 medical conditions, while AUC covers 36, excluding OASIS.

MethodACC (%)Macro-F1 (%)AUC (%)ASR (%)
No defense84.04 ± 0.1572.65 ± 0.2096.82 ± 0.0677.22 ± 0.26
NP74.18 ± 1.0254.48 ± 1.3495.31 ± 0.3322.56 ± 2.03
SAU77.60 ± 0.2459.57 ± 0.6694.12 ± 0.0612.40 ± 0.28
TSBD80.25 ± 0.0966.49 ± 0.0495.22 ± 0.1616.15 ± 0.15
TAPQ-Net87.37 ± 0.3577.23 ± 0.3298.24 ± 0.180.42 ± 0.10

Within the primary suite, TAPQ-Net leads mean ACC/Macro-F1 in 42/42 conditions and available AUC in 36/36, with lowest or tied-lowest ASR in 36/42. These are observed mean rankings (§5.1, p. 6; Appendix C, pp. 20–22).

2. What the expanded evaluation adds

The evaluation now covers additional defense families, attack constructions, selection policies and a nonmedical task. Shared partitions, poison indices, triggers and evaluation are specified alongside method-specific inputs, stages and clean-data resources (Appendix B.5, pp. 18–19).

  • Additional defenses: ABL, DBD, CBD, ASD and filtering/detection configurations are reported in Tables 3 and 15. Across the six focal dataset–mapping conditions, TAPQ-Net gives 61.21 ± 0.47% Macro-F1 and 0.61 ± 0.16% ASR (Table 3, p. 8; Table 15, p. 22).
  • Clean-data requirements: NP, SAU, TSBD and SCAn receive nominal 5% trusted subsets; TAPQ-Net receives no such subset. The primary protocol uses the fixed final checkpoint. Table 22 separately compares validation-selection policies for four baselines and three focal TAPQ-Net settings (§5.6, pp. 9–10; Appendix G, p. 25; Table 22, p. 27).
  • Extended attacks: Table 20, p. 25, covers eight constructions on three datasets and six evaluated methods/configurations: 144 cells. The joint white-box construction is an additional evaluation in Appendix E.2, p. 24.

The attack examples below are three-seed mean Macro-F1 / ASR (%) under A2O 10% poisoning:

AttackPathMNISTDermaMNISTOASIS
SIG85.73 / 0.2757.67 / 0.5539.54 / 0.31
WaNet86.88 / 0.3758.59 / 0.4038.30 / 0.36
Latent separability86.08 / 0.3658.52 / 0.5638.11 / 1.04
Joint white-box86.08 / 0.4858.52 / 0.6438.11 / 1.18

For the joint attack, undefended ASRs are 99.97/99.74/100.00%; the reported three-dataset means are 99.90% undefended and 0.77% defended. All seven gate checks still pass per seed and dataset (Table 6, p. 10; Appendix E.2, p. 24).

CIFAR-10 adds a separate ten-class A2O 10% test: 79.74% ACC, 78.85% Macro-F1 and 1.21% ASR, versus 96.51% undefended ASR (Table 24, p. 28).

3. Component attribution and remaining scope

All 126 primary runs accept all 882 eligible checks. The primary results therefore do not isolate a benefit from refresh vetoes. Table 4 uses different circuit-evaluation pathways for the full model and refresh controls; the full gate-on/off pair using one implementation is not reported (§5.3, p. 8; Table 4, p. 9; Appendix D.2, p. 23).

Classical encoders remain competitive, and the TQE has an exact classical realization (§5.4, pp. 8–9; Appendix A.1, p. 14). The theory states conditional prototype and prediction-margin results; the shot/channel evaluations are simulations (Appendix A, pp. 14–16; Appendix F, p. 25; Figure 6, p. 27). Table 23 averages each dataset's six conditions within seed, then gives descriptive intervals from all 27 ordered three-seed bootstrap resamples; these do not estimate patient-level uncertainty (Appendix H, p. 26; Table 23, p. 27).

The revised contribution is the specified prototype-guided training configuration, its characterization and its evaluated utility/ASR trade-offs. We respectfully ask that this experimental record and stated scope be considered in your final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.