Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Proceedings of the AAAI Symposium Series

Auditing Hybrid Quantum Machine Unlearning for Membership Attacks

Abhiram Manuguri, Tanish Mittal, Theshani Nuradha Piliththuwasam Gallage

Published Nov 6, 2026• 0 citations
Publisher

Abstract

Some machine learning models are known to memorize their training data, exposing them to membership inference at- tacks (MIAs) that infer whether a particular data sample was used during training. Such leakage poses serious privacy risks in sensitive domains such as security and healthcare, where membership alone can reveal protected information about an individual. Machine unlearning has been proposed as a remedy, allowing a model to “forget” withdrawn data with- out retraining from scratch. There is some evidence from re- cent works that quantum architectures such as hybrid quan- tum neural networks (HQNNs) along with unlearning meth- ods may be robust to MIAs. In this work, we audit gradient- ascent-based unlearning in HQNNs to assess how effectively forgetting happens when the model is subjected to a range of membership inference attacks. By training an HQNN that combines classical convolutional pre-processing, amplitude encoding, and a parameterized quantum circuit, and apply- ing class-level unlearning to forget a target class, we evalu- ate the unlearned model against metric-based attacks (loss, confidence, entropy, and modified entropy), and quantum at- tacks constructed from a variational quantum circuit and a quantum support vector machine. Our results show that the apparent success of unlearning is highly metric- and attack- dependent. In particular, fixed-threshold-based metric attacks seem to classify almost all forgotten class samples as non- members after unlearning. However, class-conditional area under the curve (AUC) remains above random guessing and relatively higher compared to the original model, indicating that the membership signal does not completely disappear for both the forgotten and retained classes. We observe a similar behavior with quantum attack models, even though the strength of the attacks is relatively weaker compared to metric-based ones. These findings highlight the importance of evaluating quantum unlearning against a diverse set of attacks and performance measures before concluding that privacy is preserved, and lay a foundation for systematically auditing unlearning methods in quantum machine learning models.

DOI
10.1609/aaaiss.v10i1.43491

Related papers

From membership-privacy leakage to quantum machine unlearning

Junjian Su, Runze He, Guanghui Li +4 more

Quantum machine learning (QML) has the potential to achieve quantum advantage for specific tasks by combining quantum computation with classical machine learning (ML). In classical ML, a significant challenge is membership-privacy leakage, whereby an attacker can infer from model outputs whether specific data were used in training. When specific data are required to be withdrawn, removing their influence from the trained model becomes necessary. Machine unlearning (MU) addresses this issue by enabling the model to forget the withdrawn data, thereby preventing membership-privacy leakage. However, this leakage remains underexplored in QML. This raises two research questions: do QML models leak membership privacy about their training data, and can MU methods efficiently mitigate such leakage in QML models? We investigate these questions using two quantum neural network (QNN) architectures, a basic QNN and a hybrid QNN, evaluated in noiseless simulations and cloud quantum device demonstrations. To answer the first question, we analyze how quantum constraints shape membership-privacy leakage in QML and then formalize a realistic gray-box threat model accordingly. Based on this, we design a membership inference attack (MIA) tailored to QNN outputs, and our results provide clear evidence of membership leakage in both QNNs. To answer the second question, we propose a quantum machine unlearning (QMU) framework, comprising three MU mechanisms. Evaluations on two QNN architectures show that QMU removes the influence of the withdrawn data while preserving accuracy for retained data. A comparative analysis further characterizes the three MU mechanisms with respect to data dependence, computational cost, and robustness.

arXiv:2509.06086•Physical Review Applied•Mar 23, 2026•4 citations

Distribution-Guided and Constrained Quantum Machine Unlearning

Nausherwan Malik, Zubair Khalid, Muhammad Faryad

Machine unlearning aims to remove the influence of specific training data from a learned model without full retraining. While recent work has begun to explore unlearning in quantum machine learning, existing approaches largely rely on fixed, uniform target distributions and do not explicitly control the trade-off between forgetting and retained model behaviour. In this work, we propose a distribution-guided framework for class-level quantum machine unlearning that treats unlearning as a constrained optimization problem. Our method introduces a tunable target distribution derived from model similarity statistics, decoupling the suppression of forgotten-class confidence from assumptions about redistribution among retained classes. We further incorporate an anchor-based preservation constraint that explicitly maintains predictive behaviour on selected retained data, yielding a controlled optimization trajectory that limits deviation from the original model. We evaluate the approach on variational quantum classifiers trained on the Iris and Covertype datasets. Results demonstrate sharp suppression of forgotten-class confidence, minimal degradation of retained-class performance, and closer alignment with the gold retrained model baselines compared to uniform-target unlearning. These findings highlight the importance of target design and constraint-based formulations for reliable and interpretable quantum machine unlearning.

cs.LGCross-listedarXiv:2601.04413•Jan 9, 2026•0 citations

HQSU: Hybrid Quantum Selective Unlearning for Backdoor Mitigation in Federated Learning

Asitha Kottahachchi Kankanamge Don, Ibrahim Khalil, Shehan Edirimannage +3 more

Federated Learning (FL) enables privacy-preserving collaborative training but faces significant challenges when client data must be removed or malicious updates forgotten. Regulatory requirements and backdoor threats necessitate efficient, selective unlearning methods. However, existing approximate approaches modify large portions of model parameters, increasing vulnerability to gradient inversion attacks and degrading utility. We propose HQSU (Hybrid Quantum Selective Unlearning), a hybrid quantum-classical framework that performs selective unlearning via low-rank Hessian subspace modeling and variational quantum influence estimation. HQSU preserves 96.83% clean accuracy and reduces backdoor attack success to 9.95%, matching full retraining while modifying only 0.314% of parameters, over$300 \times$fewer than dense-update baselines. Experiments on MNIST show that HQSU achieves the lowest residual backdoor accuracy among approximate methods with stable recovery dynamics. Circuit ablations indicate that amplitude-encoded, circularly entangled ansätze provide the strongest influence estimation under NISQ constraints. These results demonstrate that hybrid quantum modeling can significantly improve selectivity, stability, and privacy in federated unlearning.

International Conference on Quantum Communications, Networking, and Computing (QCNC)•Apr 6, 2026•0 citations

SQU: Sensitivity-Guided Quantum Machine Unlearning for Privacy-Preserving ML

Md Akmol Masud, Zannat Hossain Tamim, MD Musa Kalimullah Ratul +1 more

The Right to Be Forgotten (RTBF), as set out in regulations such as the General Data Protection Regulation (GDPR), mandates mechanisms to remove specific data points from trained machine learning models. Machine unlearning is a concept in deep learning that has received significant attention but remains underexplored in Quantum Machine Learning (QML). This paper presents a new model for adaptive data deletion in Variational Quantum Circuits (VQCs), called Sensitivity-Guided Quantum Unlearning (SQU). SQU combines gradient-based sensitivity analysis with a dynamic allocation of privacy budgets and targeted noise injection to eliminate the impact of high-impact samples, then fine-tunes the model on the remaining data. Experiments on three benchmark datasets, Breast Cancer, MNIST, and Fashion-MNIST, indicate that SQU successfully unlearns specified samples without diminishing the model's utility. On the Breast Cancer dataset, SQU matches the accuracy of the gold-standard retrain-from-scratch method (84.4%) while providing a quantifiable and significant privacy benefit, reducing the Membership Inference Attack (MIA) AUC by 0.67%. MNIST and Fashion-MNIST results also indicate that SQU has high utility (97.9% and 79.2%, respectively), consistent with the retraining results. A comparison with a classical neural network ($200 \times$more parameters) further demonstrates that the VQC's compact parameterization enables an exact, complete sensitivity analysis for unlearning — an advantage that is infeasible at the classical scale. In general, SQU provides a computationally efficient, regulatory-compliant solution for privacy-aware QML. Code is available at https://github.com/masud1901/Sensitivity-Guided-Quantum-Machine-Unlearning.

IEEE International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN)•Apr 16, 2026•0 citations

Self-revealing poisons: loss-guided forensic detection and quantum unlearning of corrupted training data

Oum Gadani, Kandarp Gajjar, Himani Trivedi +1 more

Abstract Quantum Computation has entered the Noisy Intermediate State Quantum era, in which quantum machine learning (QML) models built on parameterized quantum circuits are promptly utilized for real-world classification tasks, optimization and simulations because of quantum inherent properties like superposition and entanglement. Despite having such use cases, QML model’s security under adversarial conditions remains poorly understood. One such threat is data poisoning, in which an attacker corrupts the training set before the model ever sees it. Such an attack is escalated further in quantum settings by QU antum I ndiscriminate D ata Poisoning, which exploits the geometry of the quantum feature space. To address it, this paper proposes a self-revealing defense framework built on a two-stage pipeline. The first stage identifies potential corrupted data points using unsupervised hybrid k means _ gmm threshold derived from a clean reference model, enabling reliable detection without requiring any prior knowledge of the underlying attack. The second stage uses the detected samples to guide a quantum unlearning process that removes the influence of corrupted data and restores model integrity. All experiments were carried out on simulations (no real quantum hardware is utilized). Experimentation on the MNIST-4 dataset using PQC-8 confirms the attack is self-revealing, as poisoning intensifies the detection PR-AUC climbs from 0.956 to over 0.999, while receiver operating characteristic AUC remains ⩾ 0.986 throughout, achieving a robust detection and clear separability between clean and poisoned samples. Among five evaluated methods used in the quantum unlearning phase of the framework, GA, SCRUB, and Continued Fine-Tuning (CF) recover accuracy to within 10% of the clean baseline for poison ratios up to ε ⩽ 0.5 , with CF achieving this at roughly half the computational cost of the gradient-based alternatives. Additional experiments on classical baseline further reveals the proposed framework being paradigm agnostic.

Physica Scripta•Aug 11, 2026•0 citations

Federated Unlearning of Quantum Data via Classical Shadows

Zihao Ding, Jun Huang, Cong-cong Xing

Quantum federated learning trains a quantum neural network (QNN) across nodes that hold quantum data, such as the states of quantum sensors. To forget data, unlearning methods for quantum models replay them, and federated unlearning from stored updates removes whole nodes only. Quantum data cannot be replayed, since a state is consumed when measured and cannot be cloned. We propose Shadow-based Data Erasure (SHADE), in which each node records a classical shadow of every training state, a few randomized single-qubit measurements taken at training time. The gradient and curvature of a QNN on a state are expectations of observables, so SHADE estimates them for the forget data from the shadows at any later parameters and removes their effect by a federated Newton correction. Only classical estimates cross the network. We prove that the shadow size needed for a target unlearning error grows exponentially in the size of the backward light cone of the measured observable and not in the number of qubits. In simulation with 50 nodes and on edge boards, SHADE matches the forget accuracy of retraining within 0.5 points, without replay and in under 0.2 s.

quant-pharXiv:2610.19077•Oct 15, 2026•0 citations
Some machine learning models are known to memorize their training data, exposing them to membership inference at- tacks (MIAs) that infer whether a particular data sample was used during training. Such leakage poses serious privacy risks in sensitive domains such as security and healthcare, where membership alone can reveal protected information about an individual. Machine unlearning has been proposed as a remedy, allowing a model to “forget” withdrawn data with- out retraining from scratch. There is some evidence from re- cent works that quantum architectures such as hybrid quan- tum neural networks (HQNNs) along with unlearning meth- ods may be robust to MIAs. In this work, we audit gradient- ascent-based unlearning in HQNNs to assess how effectively forgetting happens when the model is subjected to a range of membership inference attacks. By training an HQNN that combines classical convolutional pre-processing, amplitude encoding, and a parameterized quantum circuit, and apply- ing class-level unlearning to forget a target class, we evalu- ate the unlearned model against metric-based attacks (loss, confidence, entropy, and modified entropy), and quantum at- tacks constructed from a variational quantum circuit and a quantum support vector machine. Our results show that the apparent success of unlearning is highly metric- and attack- dependent. In particular, fixed-threshold-based metric attacks seem to classify almost all forgotten class samples as non- members after unlearning. However, class-conditional area under the curve (AUC) remains above random guessing and relatively higher compared to the original model, indicating that the membership signal does not completely disappear for both the forgotten and retained classes. We observe a similar behavior with quantum attack models, even though the strength of the attacks is relatively weaker compared to metric-based ones. These findings highlight the importance of evaluating quantum unlearning against a diverse set of attacks and performance measures before concluding that privacy is preserved, and lay a foundation for systematically auditing unlearning methods in quantum machine learning models.
Quantum machine learning (QML) has the potential to achieve quantum advantage for specific tasks by combining quantum computation with classical machine learning (ML). In classical ML, a significant challenge is membership-privacy leakage, whereby an attacker can infer from model outputs whether specific data were used in training. When specific data are required to be withdrawn, removing their influence from the trained model becomes necessary. Machine unlearning (MU) addresses this issue by enabling the model to forget the withdrawn data, thereby preventing membership-privacy leakage. However, this leakage remains underexplored in QML. This raises two research questions: do QML models leak membership privacy about their training data, and can MU methods efficiently mitigate such leakage in QML models? We investigate these questions using two quantum neural network (QNN) architectures, a basic QNN and a hybrid QNN, evaluated in noiseless simulations and cloud quantum device demonstrations. To answer the first question, we analyze how quantum constraints shape membership-privacy leakage in QML and then formalize a realistic gray-box threat model accordingly. Based on this, we design a membership inference attack (MIA) tailored to QNN outputs, and our results provide clear evidence of membership leakage in both QNNs. To answer the second question, we propose a quantum machine unlearning (QMU) framework, comprising three MU mechanisms. Evaluations on two QNN architectures show that QMU removes the influence of the withdrawn data while preserving accuracy for retained data. A comparative analysis further characterizes the three MU mechanisms with respect to data dependence, computational cost, and robustness.
Machine unlearning aims to remove the influence of specific training data from a learned model without full retraining. While recent work has begun to explore unlearning in quantum machine learning, existing approaches largely rely on fixed, uniform target distributions and do not explicitly control the trade-off between forgetting and retained model behaviour. In this work, we propose a distribution-guided framework for class-level quantum machine unlearning that treats unlearning as a constrained optimization problem. Our method introduces a tunable target distribution derived from model similarity statistics, decoupling the suppression of forgotten-class confidence from assumptions about redistribution among retained classes. We further incorporate an anchor-based preservation constraint that explicitly maintains predictive behaviour on selected retained data, yielding a controlled optimization trajectory that limits deviation from the original model. We evaluate the approach on variational quantum classifiers trained on the Iris and Covertype datasets. Results demonstrate sharp suppression of forgotten-class confidence, minimal degradation of retained-class performance, and closer alignment with the gold retrained model baselines compared to uniform-target unlearning. These findings highlight the importance of target design and constraint-based formulations for reliable and interpretable quantum machine unlearning.
Federated Learning (FL) enables privacy-preserving collaborative training but faces significant challenges when client data must be removed or malicious updates forgotten. Regulatory requirements and backdoor threats necessitate efficient, selective unlearning methods. However, existing approximate approaches modify large portions of model parameters, increasing vulnerability to gradient inversion attacks and degrading utility. We propose HQSU (Hybrid Quantum Selective Unlearning), a hybrid quantum-classical framework that performs selective unlearning via low-rank Hessian subspace modeling and variational quantum influence estimation. HQSU preserves 96.83% clean accuracy and reduces backdoor attack success to 9.95%, matching full retraining while modifying only 0.314% of parameters, over300×300 \times300×fewer than dense-update baselines. Experiments on MNIST show that HQSU achieves the lowest residual backdoor accuracy among approximate methods with stable recovery dynamics. Circuit ablations indicate that amplitude-encoded, circularly entangled ansätze provide the strongest influence estimation under NISQ constraints. These results demonstrate that hybrid quantum modeling can significantly improve selectivity, stability, and privacy in federated unlearning.
The Right to Be Forgotten (RTBF), as set out in regulations such as the General Data Protection Regulation (GDPR), mandates mechanisms to remove specific data points from trained machine learning models. Machine unlearning is a concept in deep learning that has received significant attention but remains underexplored in Quantum Machine Learning (QML). This paper presents a new model for adaptive data deletion in Variational Quantum Circuits (VQCs), called Sensitivity-Guided Quantum Unlearning (SQU). SQU combines gradient-based sensitivity analysis with a dynamic allocation of privacy budgets and targeted noise injection to eliminate the impact of high-impact samples, then fine-tunes the model on the remaining data. Experiments on three benchmark datasets, Breast Cancer, MNIST, and Fashion-MNIST, indicate that SQU successfully unlearns specified samples without diminishing the model's utility. On the Breast Cancer dataset, SQU matches the accuracy of the gold-standard retrain-from-scratch method (84.4%) while providing a quantifiable and significant privacy benefit, reducing the Membership Inference Attack (MIA) AUC by 0.67%. MNIST and Fashion-MNIST results also indicate that SQU has high utility (97.9% and 79.2%, respectively), consistent with the retraining results. A comparison with a classical neural network (200×200 \times200×more parameters) further demonstrates that the VQC's compact parameterization enables an exact, complete sensitivity analysis for unlearning — an advantage that is infeasible at the classical scale. In general, SQU provides a computationally efficient, regulatory-compliant solution for privacy-aware QML. Code is available at https://github.com/masud1901/Sensitivity-Guided-Quantum-Machine-Unlearning.
Abstract Quantum Computation has entered the Noisy Intermediate State Quantum era, in which quantum machine learning (QML) models built on parameterized quantum circuits are promptly utilized for real-world classification tasks, optimization and simulations because of quantum inherent properties like superposition and entanglement. Despite having such use cases, QML model’s security under adversarial conditions remains poorly understood. One such threat is data poisoning, in which an attacker corrupts the training set before the model ever sees it. Such an attack is escalated further in quantum settings by QU antum I ndiscriminate D ata Poisoning, which exploits the geometry of the quantum feature space. To address it, this paper proposes a self-revealing defense framework built on a two-stage pipeline. The first stage identifies potential corrupted data points using unsupervised hybrid k means _ gmm threshold derived from a clean reference model, enabling reliable detection without requiring any prior knowledge of the underlying attack. The second stage uses the detected samples to guide a quantum unlearning process that removes the influence of corrupted data and restores model integrity. All experiments were carried out on simulations (no real quantum hardware is utilized). Experimentation on the MNIST-4 dataset using PQC-8 confirms the attack is self-revealing, as poisoning intensifies the detection PR-AUC climbs from 0.956 to over 0.999, while receiver operating characteristic AUC remains ⩾ 0.986 throughout, achieving a robust detection and clear separability between clean and poisoned samples. Among five evaluated methods used in the quantum unlearning phase of the framework, GA, SCRUB, and Continued Fine-Tuning (CF) recover accuracy to within 10% of the clean baseline for poison ratios up to ε ⩽ 0.5 , with CF achieving this at roughly half the computational cost of the gradient-based alternatives. Additional experiments on classical baseline further reveals the proposed framework being paradigm agnostic.
Quantum federated learning trains a quantum neural network (QNN) across nodes that hold quantum data, such as the states of quantum sensors. To forget data, unlearning methods for quantum models replay them, and federated unlearning from stored updates removes whole nodes only. Quantum data cannot be replayed, since a state is consumed when measured and cannot be cloned. We propose Shadow-based Data Erasure (SHADE), in which each node records a classical shadow of every training state, a few randomized single-qubit measurements taken at training time. The gradient and curvature of a QNN on a state are expectations of observables, so SHADE estimates them for the forget data from the shadows at any later parameters and removes their effect by a federated Newton correction. Only classical estimates cross the network. We prove that the shadow size needed for a target unlearning error grows exponentially in the size of the backward light cone of the measured observable and not in the number of qubits. In simulation with 50 nodes and on edge boards, SHADE matches the forget accuracy of retraining within 0.5 points, without replay and in under 0.2 s.