| name | neural-variability-enhances-robustness |
| description | 神经变异性增强人工神经网络鲁棒性方法论。研究相关性噪声如何改善对抗攻击和自然图像修改的鲁棒性，建立生物学可解释的鲁棒神经网络设计策略。 |
| author | Robin Preble, Praveen Venkatesh, Stefan Mihalas, Kameron Decker Harris |
| arxiv_id | 2606.13801 |
| categories | ["neuroscience","machine-learning","robustness","adversarial-defense"] |
| tags | ["neural-variability","noise-correlations","adversarial-robustness","naturalistic-modifications","structured-noise","local-information","biological-plausibility"] |
| created | 2026-06-15T00:00:00.000Z |
| source | arXiv cs.LG/q-bio.NC |
Neural Variability Enhances Artificial Network Robustness
Overview
This paper investigates how correlated noise in neural activations can enhance artificial neural network robustness to adversarial attacks and naturalistic image modifications. The key insight is that structured variability—similar to trial-to-trial variability observed in biological cortex—provides a biologically plausible defense mechanism using only local information.
Key Contributions
1. Biological Motivation → Computational Strategy
- Cortex variability: Trial-to-trial variability in cortical responses vs. peripheral sensory neuron consistency
- Biological hypothesis: Stochasticity may carry functional meaning
- Translation to AI: Structured noise improves network robustness
2. Noise Structure Transfer Analysis
Naturalistic Modifications:
- Structure from naturalistic modifications benefits most for similar modifications
- Poor transfer: Structure doesn't generalize across modification types
Adversarial Attacks:
- Noise structure from adversarial attacks generalizes to other attack types
- Cross-attack robustness transfer observed
3. Covariance-Based Robustness Mechanism
- Use activation covariance under modified vs. clean inputs
- Structured noise patterns derived from real-world modifications
- Local information only → biological plausibility
Core Methodology
Noise Correlation Framework
Components:
- Trial-to-trial variability measurement
- Noise-signal correlation analysis
- Covariance estimation under input perturbations
- Transfer evaluation across perturbation types
Key Metrics:
- Robustness to adversarial attacks
- Robustness to naturalistic modifications
- Cross-modification/attack transfer
Mathematical Model
Activation covariance analysis:
- Measure $C_{modified}$ - covariance under perturbed inputs
- Measure $C_{clean}$ - covariance under clean inputs
- Structure difference: $\Delta C = C_{modified} - C_{clean}$
- Robustness correlation with noise structure
Structured noise injection:
- Add noise with covariance matching $C_{modified}$
- Evaluate performance under attacks/modifications
- Compare against unstructured noise baseline
