Quantum-enhanced adversarial sample detection for robust deep learning security models
Adversarial samples remain one of the most practical threats to deep learning systems because they exploit a quiet mismatch between machine perception and human perception. A model may preserve high accuracy on clean benchmarks, yet a small and deliberately shaped perturbation can push its decision boundary into the wrong region. This paper studies a quantum-enhanced adversarial sample detection framework for robust deep learning security models. The central idea is not to replace the classical classifier, but to add a second analytic view of the input through a compact latent projection, a quantum feature map, measurement statistics, and calibrated risk fusion. The detector combines conventional security indicators such as confidence, entropy, margin, reconstruction residual, and local density with quantum-inspired kernel similarity, measurement purity, and re-encoding consistency. The resulting design treats adversarial detection as a risk estimation problem rather than a binary trick that can be solved by one score. To ground the discussion, a reproducible MNIST case study is included. Using 12000 training samples and 2500 test samples, a lightweight classifier achieved 0.909 clean accuracy and 0.079 accuracy under an FGSM-style attack with epsilon 0.06. A classical detector reached AUC 0.887, while the quantum-enhanced fusion branch reached AUC 0.959. The results are illustrative rather than a final benchmark, but they show why quantum feature spaces are attractive for security monitoring: they can expose distributional changes that are not fully visible in logits alone. To address the reviewers’ concerns on generalization and rigor, the framework is further evaluated on CIFAR-10 with a convolutional classifier and under stronger attacks (PGD, CW, DeepFool, AutoAttack, spatial-transformation, and patch attacks), and is complemented by per-feature ablations, calibration metrics (ECE and Brier score), computational-cost analysis, and comparisons against strong classical detectors and equally expressive classical kernels. We also clarify that the quantum branch is realized here by a classical, quantum-inspired emulation.
Authors
- XinHang He
Institutions
- Beihang University (CN)
Publication Details
- Journal
- Discover Computing
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1007/s10791-026-10849-9
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00