| name | sqnn-decoherence-defence |
| description | Stochastic Quantum Neural Networks (SQNNs) adversarial robustness methodology. Decoherence-contraction theorem for depolarising channels, per-gate dropout as curvature-weighted L2 penalty, and noise-as-defence framework for quantum machine learning. Use when: SQNNs, quantum adversarial robustness, quantum dropout, decoherence contraction, Lindblad master equation, neural network security. |
| metadata | {"arxiv_id":"2606.24219","published":"2026-06-23","authors":"Gautier-Edouard Filardo","tags":["quantum","machine-learning","adversarial-robustness","sqnn","decoherence"]} |
SQNN Decoherence as Defence
Description
Stochastic Quantum Neural Networks (SQNNs) encode neuronal activations as qubits, synaptic topology as entanglement, and neural noise through a Lindblad master equation. This methodology proves that decoherence acts as adversarial defence rather than degradation.
Activation Keywords
- sqnn decoherence defence
- stochastic quantum neural network robustness
- quantum dropout
- decoherence contraction theorem
- depolarising channel regularisation
- quantum adversarial robustness
- lindblad master equation neural
- per-gate stochastic deactivation
Core Contributions
1. Decoherence-Contraction Theorem
A depolarising channel of strength γ over L entangling layers contracts every weight-w Pauli read-out by factor:
(1 - 4γ/3)^(wL)
For weight-1 read-out: (1 - 4γ/3)^L
Key insight: Depolarising noise reshapes the training boundary rather than contracting attack-time gradients. This prevents catastrophic robustness collapse seen in noiseless circuits.
2. Per-Gate Dropout as Curvature-Weighted L2 Penalty
Per-gate dropout implements:
p(1-p)/2 × Σ θ² ∂²_θ L
in weight space, maximised at p=1/2. Depolarising noise implements an output-space penalty instead.
Practical result: Both mechanisms reduce train-test gap by ~0.01 (p<10⁻⁴), statistically indistinguishable. Effect concentrated where overfitting is largest.
3. Empirical Validation
- NSL-KDD dataset under FGSM and PGD attacks
- 30-seed study confirms dropout formula prediction
- Increasing dropout past 1/2 does not help
- Noiseless models fall from 95% to 47% under PGD-20; SQNN never suffers catastrophic collapse
Usage Workflow
Step 1: SQNN Architecture Design
- Encode neuronal activations as qubits
- Map synaptic topology to entanglement structure
- Choose ring topology for non-local anomaly detection
Step 2: Noise Mechanism Selection
| Mechanism | Effect | Best For |
|---|
| Depolarising (γ) | Output-space penalty | General robustness |
| Per-gate dropout (p) | Curvature-weighted L2 | Overfitting prevention |
| Combined | Both penalties | Maximum robustness |
Step 3: Training with Noise-as-Defence
- Apply depolarising channel during training (not just inference)
- Set dropout rate p ≤ 0.5 (optimal at p=1/2)
- Monitor robustness variance (should decrease ~2×)
- Validate under white-box FGSM and PGD attacks
Step 4: Neutral-Atom Realisation
For hardware deployment:
- Map SQNN to neutral-atom architecture
- Perform feasibility analysis by N qubits
- Account for native gate set and connectivity
Pitfalls
- Depolarising ≠ dropout: Depolarising channel acts as output noise, not dropout-style regulariser. Use per-gate dropout for weight-space effects.
- Optimal dropout at p=0.5: Increasing beyond 0.5 does not improve generalisation.
- Single-seed results unreliable: Prior work with single seeds showed false dichotomy; use 30+ seeds.
- Robustness from boundary reshaping: Robustness comes from noise-reshaped training boundary, not gradient contraction during attack.
Related Skills
stochastic-quantum-neural-network — SQNN basics (arXiv:2511.11609)
qmt-quantum-measurement-temperature — QNN training stability (arXiv:2606.22551)
qpinn-integro-fractional-pde — QPINN methodology (arXiv:2606.26865)
qml-adversarial-robustness-sok — QML robustness SoK