Feature-level analysis and adversarial transfer in rotationally equivariant quantum machine learning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.
Kai: We're looking at the paper: "Feature-level analysis and adversarial transfer in rotationally equivariant quantum machine learning".
Paper summary: Kai: We've talked about what this paper is about—it’s not just about equivariance improving trainability; it’s specifically about how symmetry constraints shape adversarial robustness. The paper, "Feature-level analysis and adversarial transfer in rotationally equivariant quantum machine learning," studies this question through a feature-level analysis of these models in a transfer-attack setting.
Mira: It claims that under equivariance with an invariant readout, the predictions only depend on the group twirled input, which identifies the symmetry invariant information accessible to the model alongside some complementary uninformative subspace.
Lev: And they specialize this framework to rotationally equivariant quantum models and derive a specific characterization of what information is accessible in terms of rotation invariant image statistics distributed across distinct symmetry sectors.
Kai: They then use targeted input transformations to determine which of these statistics are actually being relied upon for classification across several different datasets.
Mira: The core finding is that the model only accesses rotationally invariant image statistics that decompose across its symmetry channels, and these accessible features are governed by circular correlations called C r,r prime delta phi.
Lev: To probe this, they introduce two ways to test which statistics are used: targeted input transformations and modifications at the readout level. They show that while the model is restricted to rotation invariant information, the particular invariant statistics it relies on change a lot depending on which dataset you use.
Kai: Then they look at robustness under transfer attacks and find that when trained on a transformation called T3, which removes ring averages, the model becomes substantially more robust than when trained on clean data.
Mira: This is interesting because even for standard datasets like STM or RotFMNIST, the clean-trained model already shows poor robustness, similar to what you'd see with a transformation called T2.
Lev: The authors suggest that ring-average information accounts for a large chunk of the discriminative signal these models use on these specific types of data.
Kai: So, their main takeaway is that suppressing the symmetry channel associated with those ring averaged intensities substantially improves robustness, and they find this effect is strongest on datasets where that feature was previously most vulnerable.
Conclusion: Mira: So to wrap up, the paper "Feature-level analysis and adversarial transfer in rotationally equivariant quantum machine learning" shows us that simply having equivariance isn't a complete solution for adversarial robustness.
Kai: They demonstrate that simple invariant features, like the ring-averaged intensities in this rotationally equivariant model, can remain highly predictive while being very non robust to transfer attacks.
Lev: The paper provides a mechanistic understanding of how these quantum models use their symmetry channels to make decisions, which helps explain properties like adversarial robustness.
Mira: The authors show that by performing a feature-level analysis and suppressing the specific symmetry channel responsible for those brittle features, we can substantially improve robustness without losing too much accuracy.
Kai: This opens up the idea that we can use this framework to get a mechanistic look at other architectures with different types of symmetry constraints.
Lev: It gives us a concrete way to understand exactly where the vulnerabilities are hidden in these quantum models, moving beyond just observing that they are fragile under attack.
Maureen Krumtülnger, Martin Sevior, Muhammad Usman
School of Physics, University of Melbourne · Data61, CSIRO
quant-ph
Submitted: 2026-04-16
Updated: 2026-04-16
Journal ref: Advanced Quantum Technologies (2026)
DOI: 10.1002/qute.70521
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Group-equivariant quantum models are designed to exploit symmetry and can improve trainability, but it remains unclear how symmetry constraints shape their adversarial robustness The authors study…
Key concepts
- Group-equivariant quantum models
- These are quantum machine learning models designed to respect a specific mathematical symmetry group. The goal is to use this inherent symmetry to make the model better at learning and potentially more robust against attacks. However, the study shows that just having this symmetry isn't enough on its own.
- Invariant Readout
- This means the final prediction of the model only depends on information that remains unchanged when you apply a transformation from a specific symmetry group. This limits what the model can 'see' but also defines the core, symmetric information it uses for classification.
- Brittle Statistics
- These are specific types of features, such as ring-averaged intensities in the rotationally equivariant model, that seem important for training but are not reliable when facing adversarial attacks. They are fragile; small changes to these specific statistics can cause large prediction errors under transfer attacks.
- Transfer Attacks
- These are adversarial attacks where an attacker creates an input perturbation designed to fool a model trained on one dataset, and that same attack successfully fools the model when tested on a different, unseen dataset. The paper examines how symmetry affects the success of these attacks.
Terminology
Summary
Group-equivariant quantum models are designed to exploit symmetry and can improve trainability, but it remains unclear how symmetry constraints shape their adversarial robustness The authors study this question through a feature-level analysis of equivariant quantum models in a transfer-attack setting. p. 0 Feature-level analysis and adversarial transfer in rotationally equivariant
quantum machine learning
Maureen Krumtünger,1, ∗ Martin Sevior,1
and Muhammad Usman1, 2
1School of Physics, University of Melbourne, Parkville, VIC 3010, Australia
2Data61, CSIRO, Clayton, 3168, Victoria, Australia
Group-equivariant quantum models are designed to exploit symmetry and can improve trainability, but it remains
unclear how symmetry constraints shape their adversarial robustness. We study this question through a featurelevel analysis of equivariant quantum models in a transfer-attack setting. We study this question through a featurelevel analysis of equivariant quantum models in a transfer-attack setting. Under equivariance with an invariant
readout, predictions depend only on the group-twirled input, which identifies the symmetry-invariant information
accessible to the model together with a complementary uninformative subspace. Using targeted input
transformations, we determine which of these statistics are actually relied upon for classification across several
datasets. We find that equivariance alone does not guarantee transfer robustness: even within the restricted
invariant feature space, the model can rely on brittle statistics, particularly ring-averaged intensities in the
rotationally equivariant model, that remain vulnerable to classical transfer attacks. We find that equivariance alone does not guarantee transfer robustness: even within the restricted
invariant feature space, the model can rely on brittle statistics, particularly ring-averaged intensities in the
rotationally equivariant model, that remain vulnerable to classical transfer attacks. Guided by this analysis, we
show that suppressing the symmetry sector associated with the brittle feature substantially improves robustness.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Feature Accessibility and Characterization
The model's prediction depends only on the group-twirled input, which identifies the symmetry-invariant information accessible to the model together with a complementary uninformative subspace. Specializing this framework to a rotationally equivariant quantum model, we derive an explicit characterization of the accessible information in terms of rotation-invariant image statistics distributed across distinct symmetry sectors. p. 0 Under equivariance with an invariant
readout, predictions depend only on the group-twirled input, which identifies the symmetry-invariant information
accessible to the model together with a complementary uninformative subspace. Specializing this framework to
a rotationally equivariant quantum model, we derive an explicit characterization of the accessible information
in terms of rotation-invariant image statistics distributed across distinct symmetry sectors. Specializing this framework to
a rotationally equivariant quantum model, we derive an explicit characterization of the accessible information
in terms of rotation-invariant image statistics distributed across distinct symmetry sectors. Using targeted input
transformations, we determine which of these statistics are actually relied upon for classification across several
datasets. [5] introduced this architecture and established its trainability, our contribution is different in scope:
we derive a twirling-based characterization of the input information the model can access, thereby obtaining a much clearer
understanding of how the model works. We show that the
model only accesses rotationally invariant image statistics that
decompose across its symmetry channels.Where the paper says it
Where the paper says it
Where the paper says it
Probing and Identifying Brittle Features
To probe which accessible statistics are used by the trained model, we introduce two complementary approaches to probe which accessible statistics are used by the trained model: (i) targeted input transformations and (ii) readout-level modifications. Transformation 1 (T1) is designed to directly test this prediction. p. 3 The characterization above yields a concrete prediction: a
Z𝑁𝜙
-equivariant model should be invariant under any input
transformation that leaves the twirled representation TZ𝑁𝜙
(𝜌)
unchanged. Transformation 1 (T1) is designed to directly
test this prediction. We
subtract the ring mean from each ring,
𝑥
(3)
𝑟 = 𝑥𝑟 − 𝑥¯𝑟 1, 𝑥¯𝑟 =
1
𝑁𝜙
∑︁
𝜙
𝑥𝑟, 𝜙, (17)
where 1 is an all-ones vector. This eliminates the ring averages while retaining the correlation structure up to a Δ𝜙-
independent offset (see Appendix C). At the
same time, the relatively weak performance even on T3 indicates that these features alone do not fully account for the
discriminative signal present in the clean data. Overall, while the model is restricted to rotation-invariant information, the particular invariant statistics it relies on vary
substantially across datasets.Where the paper says it
Where the paper says it
Where the paper says it
Robustness under Transfer Attacks
The model becomes substantially more robust when trained on T3 than when trained on clean data, although this comes at the cost of lower clean-test accuracy. By contrast, models trained on T2 are consistently less robust. p. 6 The model becomes substantially more robust when trained
on T3 than when trained on clean data, although this comes
at the cost of lower clean-test accuracy. By contrast, models
trained on T2 are consistently less robust. By contrast, models
trained on T2 are consistently less robust. Moreover, for STM,
RotFMNIST, and CIFAR, the clean-trained model already exhibits poor robustness, at a level similar to T2. Moreover, models trained on T2 generalize
to clean test data slightly better than T3-trained models, and
substantially better in the case of STM. Together, these results
suggest that ring-mean information accounts for a large fraction of the discriminative signal used by the model on these
datasets.Where the paper says it
Where the paper says it
Where the paper says it
Improving Robustness through Suppression
The strongest improvement is observed for the datasets that were previously most vulnerable to the ring-average feature. Projecting out the m = 0 mode, which carries the ring-averaged intensities, increases the robustness of the model without any substantial loss in accuracy. p. 8 At the
same time, the equivariant structure—especially through its irreducible decomposition—provides a principled and targeted
handle for controlling specific invariant features. In particular,
we showed that projecting out the 𝑚 = 0 sector, which carries
the ring-averaged intensities, increases the robustness of the
model without any substantial loss in accuracy. tectural intervention: training on clean data while projecting
out the 𝑚 = 0 channel, corresponding to the ring-averaged
intensities, which Section III identified as a likely source of
non-robustness. This second strategy is particularly motivated
by the fact that training on T3 also improves robustness, but at
a substantial cost in clean-test accuracy; see Table I (b). For CNN- and ResNet18-based attacks,
however, it yields a larger robustness improvement than adversarial training on the LC-generated examples. Overall, this
makes the 𝑚 = 0 projection a particularly effective method in
our setting.Where the paper says it
Where the paper says it
Where the paper says it
Implications and Outlook
Our results suggest that this is not the case in general, although equivariance renders the model insensitive to a large non-invariant subspace of perturbations. In particular, we find that simple invariant features, such as the ring-averaged intensities of the rotationally equivariant model, can remain highly predictive while being strongly non-robust. p. 8 As a result, brittle and robust features can still coexist within the restricted
class of rotation-invariant statistics. In particular, we find
that simple invariant features, such as the ring-averaged intensities of the rotationally equivariant model, can remain
highly predictive while being strongly non-robust, whereas
richer invariant structures, such as higher-order angular correlation statistics, can be substantially more robust. In particular,
we showed that projecting out the 𝑚 = 0 sector, which carries
the ring-averaged intensities, increases the robustness of the
model without any substantial loss in accuracy. More broadly, this idea could be explored further by investigating in greater detail the signal contribution of the remaining
symmetry channels, and by studying architectural strategies
that either preserve them as separate feature streams or recombine them only at a later stage. Thus, this framework could be similarly used for a mechanistic
analysis of other equivariant architectures. Overall, our study shows how equivariance can enable a mechanistic, feature-level understanding of quantum models and
thereby help explain properties such as adversarial robustness. [5], the same twirling-based reasoning extends
to other strictly equivariant quantum models; what becomes
case-specific is the interpretation of the resulting symmetry
channels in terms of input-space features for a given encoding. Thus, this framework could be similarly used for a mechanistic
analysis of other equivariant architectures.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
[1] M. Larocca, F. p. 9 The authors acknowledge the computational resources provided by the University of Melbourne’s Research Computing Service. Larocca, F. Larocca, F. Sauvage, F. Sauvage, F. Coles, and
M.Where the paper says it
Where the paper says it
Where the paper says it
[2] J. J. Meyer, M. Mularski, E. p. 9 Cerezo, Group-invariant quantum machine learning, PRX
Quantum 3, 030341 (2022). Mularski, E. Mularski, E. Gil-Fuster, A. Gil-Fuster, A. Wilms, and J.Where the paper says it
Where the paper says it
Where the paper says it
[3] Q. T. Nguyen, L. Schatzki, P. p. 9 Eisert, Exploiting symmetry in variational
quantum machine learning, PRX Quantum 4, 010328 (2023). Schatzki, P. Schatzki, P. Braccia, M. Larocca, F. Sauvage, F. The authors acknowledge the computational resources provided by the University of Melbourne’s Research Computing Service. Larocca, F.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[4] L. Schatzki, M. p. 9 Eisert, Exploiting symmetry in variational
quantum machine learning, PRX Quantum 4, 010328 (2023). Schatzki, P. The authors acknowledge the computational resources provided by the University of Melbourne’s Research Computing Service. Larocca, F. Larocca, F. Sauvage, F. Sauvage, and
M. Cerezo, Theoretical guarantees for permutation-equivariant
quantum neural networks, npj Quantum Information 10,
10.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[5] M. T. West, J. Heredge, M. p. 9 1038/s41534-024-00804-1 (2024). Heredge, M. Feature-level analysis and adversarial transfer in rotationally equivariant
quantum machine learning
Maureen Krumtünger,1, ∗ Martin Sevior,1
and Muhammad Usman1, 2
1School of Physics, University of Melbourne, Parkville, VIC 3010, Australia
2Data61, CSIRO, Clayton, 3168, Victoria, Australia
Group-equivariant quantum models are designed to exploit symmetry and can improve trainability, but it remains
unclear how symmetry constraints shape their adversarial robustness. We study this question through a featurelevel analysis of equivariant quantum models in a transfer-attack setting. Feature-level analysis and adversarial transfer in rotationally equivariant
quantum machine learning
Maureen Krumtünger,1, ∗ Martin Sevior,1
and Muhammad Usman1, 2
1School of Physics, University of Melbourne, Parkville, VIC 3010, Australia
2Data61, CSIRO, Clayton, 3168, Victoria, Australia
Group-equivariant quantum models are designed to exploit symmetry and can improve trainability, but it remains
unclear how symmetry constraints shape their adversarial robustness. We study this question through a featurelevel analysis of equivariant quantum models in a transfer-attack setting.Where the paper says it
Where the paper says it
Where the paper says it
[6] M. T West, M. Sevior, and M. Usman, Reflection equivariant quantum neural networks for enhanced image classification, Machine Learning: Science and Technology 4, 035027 (2023)
[7] X. Yuan, P. He, Q. Zhu, and X. p. 9 Usman, Reflection equivariant quantum neural networks for enhanced image classification,
Machine Learning: Science and Technology 4, 035027 (2023). Zhu, and X. Zhu, and X. Li, Adversarial examples: Attacks and defenses for deep learning (2018), arXiv:1712.Where the paper says it
Where the paper says it
[8] O. Ibitoye, R. p. 9 Li, Adversarial examples: Attacks and defenses for deep learning (2018), arXiv:1712. Ibitoye, R. Ibitoye, R. Abou-Khamis, M. Abou-Khamis, M. el Shehaby, A. el Shehaby, A. Matrawy, and
M. Matrawy, and
M. Shafiq, The threat of adversarial attacks on machine learning in network security – a survey (2023), arXiv:1911.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[9] I. J. Goodfellow, J. p. 9 Shafiq, The threat of adversarial attacks on machine learning in network security – a survey (2023), arXiv:1911. Goodfellow, J. Goodfellow, J. Shlens, and C. Shlens, and C. Szegedy, Explaining
and harnessing adversarial examples (2015), arXiv:1412. Szegedy, Explaining
and harnessing adversarial examples (2015), arXiv:1412. 6572
[stat.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[10] D. Tsipras, S. p. 9 6572
[stat. Tsipras, S. Tsipras, S. Santurkar, L. Santurkar, L. Engstrom, A. Engstrom, A. Turner, and A. Turner, and A. Madry,
Robustness may be at odds with accuracy, in International Conference on Learning Representations (2019).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[11] A. Ilyas, S. Santurkar, D. p. 9 Tsipras, S. Santurkar, L. 6572
[stat. Tsipras, S. Santurkar, L. Engstrom, A. Engstrom, B. Tran, and
A. Tran, and
A. Madry, Adversarial examples are not bugs, they are features
(2019), arXiv:1905. Madry, Adversarial examples are not bugs, they are features
(2019), arXiv:1905. 02175 [stat.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[12] A. Demontis, M. p. 9 02175 [stat. Demontis, M. Demontis, M. Jagielski, B. Jagielski, B. Nita-Rotaru, and F. Nita-Rotaru, and F. Roli, Why do adversarial attacks transfer? Roli, Why do adversarial attacks transfer? explaining transferability of evasion and poisoning attacks (2019), arXiv:1809.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[13] A. Madry, A. Makelov, L. p. 9 explaining transferability of evasion and poisoning attacks (2019), arXiv:1809. Makelov, L. Makelov, L. Schmidt, D. 6572
[stat. Tsipras, S. Tsipras, and A. Vladu,
Towards deep learning models resistant to adversarial attacks, in
International Conference on Learning Representations (2018).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[14] M. Usman, Quantum Robustness in Artificial Intelligence: Principles and Applications, Quantum Science and Technology (Springer Nature Switzerland, 2026)
[15] S. Lu, L.-M. Duan, and D. p. 9 Usman, Quantum Robustness in Artificial Intelligence: Principles and Applications, Quantum Science and Technology
(Springer Nature Switzerland, 2026). Duan, and D. Usman, Quantum Robustness in Artificial Intelligence: Principles and Applications, Quantum Science and Technology
(Springer Nature Switzerland, 2026). Duan, and D. Duan, and D. Deng, Quantum adversarial machine learning, Physical Review Research 2, 10. Deng, Quantum adversarial machine learning, Physical Review Research 2, 10. 1103/physrevresearch. Deng, Quantum adversarial machine learning, Physical Review Research 2, 10. 1103/physrevresearch.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[16] W. Gong and D. p. 9 033212 (2020). Gong and D. Gong and D. Deng, Universal adversarial examples and
perturbations for quantum classifiers, National Science Review
10.Where the paper says it
Where the paper says it
[17] M. T. West, S. M. Erfani, C. Leckie, M. Sevior, L. C. L. Hollenberg, and M. p. 9 1093/nsr/nwab130 (2021). Hollenberg, and M. Hollenberg, and M. Usman, Benchmarking adversarially robust quantum machine learning at scale, Physical Review Research 5,
10.Where the paper says it
Where the paper says it
[18] M. T. West, S.-L. Tsang, J. S. Low, C. D Hill, C Leckie, L C L Hollenberg, S M Erfani, and M Usman, Towards quantum enhanced adversarial robustness in machine learning (Nature Machine Intelligence 5)
[19] N. Dowling, M. p. 9 Usman, Towards
quantum enhanced adversarial robustness in machine learning,
Nature Machine Intelligence 5, 581–589 (2023). Dowling, M.Where the paper says it
[20] M. Ragone, P. Braccia, Q. p. 9 Schatzki, P. Braccia, M. Larocca, and M. Cerezo, Representation theory for geometric quantum machine learning (2023),
arXiv:2210.Where the paper says it
Where the paper says it
[21] Y. Lecun, L. Bottou, Y. Bengio, and P. p. 9 07980 [quant-ph]. Bengio, and P.Where the paper says it
[22] H. Xiao, K. Rasul, and R. p. 9 Haffner, Gradient-based
learning applied to document recognition, Proceedings of the
IEEE 86, 2278 (1998). Rasul, and R. Rasul, and R. Vollgraf, Fashion-mnist: a novel
image dataset for benchmarking machine learning algorithms
(2017), arXiv:1708. Rasul, and R. Vollgraf, Fashion-mnist: a novel
image dataset for benchmarking machine learning algorithms
(2017), arXiv:1708.Where the paper says it
Where the paper says it
Where the paper says it
[23] A. Krizhevsky, Learning multiple layers of features from tiny images (2009)
[24] K. He, X. Zhang, S. Ren, and J. p. 9 Krizhevsky, Learning multiple layers of features from tiny
images (2009). Ren, and J.Where the paper says it
[27] M. Zhao, L Zhang, J Ye, H Lu, B Yin, and X Wang, Adversarial training: A survey (2024), arXiv:2410. p. 9 Yin, and X. Wang, Adversarial training: A survey (2024), arXiv:2410.Where the paper says it
[28] T. Bai, J. Luo, J Zhao, B Wen, and Q Wang, Recent advances in adversarial training for adversarial robustness (2021), arXiv:2102. p. 9 Wen, and Q. Wang, Recent advances in adversarial training for adversarial robustness (2021),
arXiv:2102.Where the paper says it
[29] A. Raghunathan, S. p. 9 Wang, Recent advances in adversarial training for adversarial robustness (2021),
arXiv:2102. Raghunathan, S. Duchi, and P. Liang,
Understanding and mitigating the tradeoff between robustness
and accuracy (2020), arXiv:2002.Where the paper says it
Where the paper says it
[30] T. Pang, M. Lin, X Yang, J Zhu, and S Yan, Robustness and accuracy could be reconcilable by (proper) definition (2022), arXiv:2202. p. 9 Zhu, and S. Yan, Robustness and
accuracy could be reconcilable by (proper) definition (2022),
arXiv:2202.Where the paper says it
[31] L Wang, I. I Uddin, K Santosh, C Zhang, X Qin, and Y Zhou, Bridging symmetry and robustness: On the role of equivariance in enhancing adversarial robustness (The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025))
[32] M. Usman, Y. Z. Wong, C D Hill, et al. p. 9 Zhou,
Bridging symmetry and robustness: On the role of equivariance
in enhancing adversarial robustness, in The Thirty-ninth Annual
Conference on Neural Information Processing Systems (2025). Hill, et al.Where the paper says it
[33] M. T West and M Usman, Framework for donor-qubit spatial metrology in silicon with depths approaching the bulk limit, Phys. p. 9 West and M. Usman, Framework for donor-qubit spatial
metrology in silicon with depths approaching the bulk limit,
Phys.Where the paper says it
[34] M. Usman, B Voisin, J Salfi, S Rogge, and L C L Hollenberg, Towards visualisation of central-cell-effects in scanning tunnelling microscope images of subsurface dopant qubits in silicon (Nanoscale 9)
[35] K. Chen and L. p. 9 Hollenberg, Towards visualisation of central-cell-effects in scanning
tunnelling microscope images of subsurface dopant qubits in
silicon, Nanoscale 9, 17013–17019 (2017). Chen and L. Hollenberg, Towards visualisation of central-cell-effects in scanning
tunnelling microscope images of subsurface dopant qubits in
silicon, Nanoscale 9, 17013–17019 (2017). Chen and L. Chen and L. Liu, Privacy preserving data classification with
rotation perturbation, in Proceedings of the Fifth IEEE International Conference on Data Mining, ICDM ’05 (IEEE Computer
Society, USA, 2005) p.Where the paper says it
Where the paper says it
Where the paper says it
[36] K. Chen and L. Liu, Geometric data perturbation for privacy preserving outsourced data mining, Knowl. p. 9 Chen and L. Liu, Geometric data perturbation for privacy preserving outsourced data mining, Knowl.Where the paper says it
[37] K. Liu, H Kargupta, and J Ryan, Random projection-based multiplicative data perturbation for privacy preserving distributed data mining, IEEE Trans. p. 9 Chen and L. Liu, Privacy preserving data classification with
rotation perturbation, in Proceedings of the Fifth IEEE International Conference on Data Mining, ICDM ’05 (IEEE Computer
Society, USA, 2005) p. However, recent studies have shown that quantum models can exhibit increased resilience to transfer attacks
crafted on classical surrogate models [17–19]. One possible
explanation is that quantum and classical learners rely on different predictive features of the same data, as evidenced by
the qualitatively different perturbation patterns observed for
attacks optimized on classical versus quantum models [17].Where the paper says it
Where the paper says it
[38] C Dwork, F McSherry, K Nissim, and A Smith, Calibrating noise to sensitivity in private data analysis (Springer-Verlag, Berlin, Heidelberg, 2006) p. 265–284
[39] M T West and M Usman, Framework for atomiclevel characterisation of quantum computer arrays by machine learning (npj Computational Materials 6)
[40] L Schatzki, M Larocca, Q T Nguyen, F Sauvage, and M Cerezo, Representation theory for geometric quantum machine learning (2023), arXiv:2210.07980 [quant-ph]
[41] Y Lecun, L. Bottou, Y Bengio, and P Haffner, Gradient-based learning applied to document recognition (Proceedings of the IEEE 86)
[42] H Xiao, K Rasul, and R. Vollgraf, Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms (2017), arXiv:1708.07747 [cs.LG]
[43] A Krizhevsky, Learning multiple layers of features from tiny images (2009)
[44] K He, X. Zhang, S Ren, and J Sun, Deep residual learning for image recognition (CVPR 2016)
[47] M T West, S M Erfani, C Leckie, M Sevior, L C L Hollenberg, and M Usman, Benchmarking adversarially robust quantum machine learning at scale (Physical Review Research 5)
[48] N Dowling, M T West, A Southwell, A C Nakhl, M Sevior, M Usman, and K Modi, Adversarial robustness guarantees for quantum classifiers (npj Quantum Information 12)
[49] S Lu, L.-M. Duan, and D.-L. Deng, Quantum adversarial machine learning (Physical Review Research 2)
[50] W Gong and D.-L. Deng, Universal adversarial examples and perturbations for quantum classifiers (National Science Review)
How the training signal is built
The rotationally equivariant quantum model is trained for 10 epochs using a depth-100 circuit. All reported results are averaged over five independent training runs. p. 4 For each training-set variant, the rotationally
equivariant model is trained for 10 epochs using a depth-100
circuit. All reported results are averaged over five independent
training runs. For each dataset, we consider the
original training set (clean) together with transformed variants
obtained by applying each transformation from Sec. III B 1 to
all training inputs while leaving the labels unchanged; these
variants are denoted T1, T2, and T3. Adversarial examples. To evaluate transfer robustness,
we generate adversarial examples using FGSM and PGD with
four classical surrogate models: a linear classifier (LC), PGD is a multi-step
variant of the FGSM attack [13]. Starting from an initial
perturbation 𝑥
0
, it iteratively updates the adversarial example
as
𝑥
𝑡+1 = ΠB𝜀 (𝑥)
𝑥
𝑡 + 𝛼 sign∇𝑥𝐿(𝜽
∗
, 𝑥𝑡
, 𝑦)
, (6)
where ΠB𝜀 (𝑥) denotes the projection onto the ℓ∞-ball B𝜀 (𝑥) =
{𝑥
′
: ∥𝑥
′ − 𝑥∥∞ ≤ 𝜀} around the original input 𝑥, 𝛼 is the
step size, and 𝐿 is the loss function of the model.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Feature-level analysis for robustness assessment: The system can perform a mechanistic understanding of which input features are actually accessible to a group equivariant quantum model by deriving a twirling-based characterization of input information. This allows for targeted input transformations, such as Transformation 1, to test the model's reliance on specific rotation-invariant statistics.
-
Robustness enhancement via feature suppression: The system can improve robustness by suppressing the symmetry sector associated with brittle features. Specifically, projecting out the m equals zero Fourier mode increases robustness without a substantial loss in accuracy across all datasets when compared to adversarial training alone.
-
Dataset-dependent feature reliance mapping: The system can map which rotation-invariant statistics are utilized by a model on a dataset-specific basis. For example, it can determine that for STM, RotFMNIST, and CIFAR, the model exhibits
a strong reliance on the information retained in T2,
while for MNIST and RotMNIST, it retains more useful information under T3. -
Practical data obfuscation: The system can implement a practical obfuscation scheme using Transformation 1 for privacy-preserving data sharing. Training on scrambled images preserves the rotation-invariant correlations accessible to the model, allowing utility comparable to training on clean inputs without exposing original pixel-level content directly.
Abstract
Group-equivariant quantum models are designed to exploit symmetry and can improve trainability, but it remains unclear how symmetry constraints shape their adversarial robustness. We study this question through a feature-level analysis of equivariant quantum models in a transfer-attack setting. Under equivariance with an invariant readout, predictions depend only on the group-twirled input, which identifies the symmetry-invariant information accessible to the model together with a complementary uninformative subspace. Specializing this framework to a rotationally equivariant quantum model, we derive an explicit characterization of the accessible information in terms of rotation-invariant image statistics distributed across distinct symmetry sectors. Using targeted input transformations, we determine which of these statistics are actually relied upon for classification across several datasets. We find that equivariance alone does not guarantee transfer robustness: even within the restricted invariant feature space, the model can rely on brittle statistics, particularly ring-averaged intensities in the rotationally equivariant model, that remain vulnerable to classical transfer attacks. Guided by this analysis, we show that suppressing the symmetry sector associated with the brittle feature substantially improves robustness. These results establish a systematic mechanism to exploit symmetry-dependent features for adversarial robustness in future quantum machine learning models.
Sources
- Adversarial Examples: Attacks and Defenses for Deep Learning
- The Threat of Adversarial Attacks on Machine Learning in Network Security -- A Survey
- Explaining and Harnessing Adversarial Examples
- Adversarial Examples Are Not Bugs, They Are Features
- Why Do Adversarial Attacks Transfer? Explaining Transferability of Evasion and Poisoning Attacks
- Representation Theory for Geometric Quantum Machine Learning
- Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
- Supervised quantum machine learning models are kernel methods
- Adversarial Training: A Survey
- Recent Advances in Adversarial Training for Adversarial Robustness
- Understanding and Mitigating the Tradeoff Between Robustness and Accuracy
- Robustness and Accuracy Could Be Reconcilable by (Proper) Definition
Related papers
- Computational regimes in matrix-product-state-based quantum trajectory simulations
- Toward chemically accurate quantum simulations on quantum hardware: A quantum-HPC hybrid approach
- Transversal surface-code game with reconfigurable qubits
- Fulqrum: a generalized framework for quantum subspace diagonalization
- Flux-Activated Resonant Control of a Bosonic Quantum Memory
- Machine learning the effects of many quantum measurements