Structured Noise Boosts AI Robustness—New Study Shows Biological Strategy Works
Adding correlated noise to neural nets improves defense against adversarial attacks by up to 40%
A new arXiv paper from researchers Robin Preble, Praveen Venkatesh, Stefan Mihalas, and Kameron Decker Harris investigates whether correlated noise—similar to the trial-to-trial variability observed in cortical neurons—can improve the robustness of artificial neural networks (ANNs). While peripheral sensory neurons respond consistently, cortical responses show high variability, suggesting noise may carry functional meaning. The team hypothesized that structured noise in ANN activations could enhance resilience against adversarial attacks and naturalistic image modifications.
Using covariance analysis of activations under modified versus clean inputs, they found that adding correlated noise significantly improves robustness. Notably, noise structure from adversarial attacks generalizes well to other attack types, but structure optimized for naturalistic modifications (e.g., blur, contrast changes) transfers poorly across modification types. This indicates that different forms of noise serve distinct purposes. The approach is biologically plausible—it relies only on local information within the network, similar to how biological neurons operate. The results suggest a new, training-free strategy for hardening AI systems against both malicious and naturally occurring input variations.
- Structured (correlated) noise in ANN activations improves robustness to adversarial attacks and naturalistic image modifications
- Noise structure from adversarial attacks generalizes across attack types, but naturalistic noise structure does not transfer between modification types
- The approach uses only local information, making it biologically plausible and computationally efficient
Why It Matters
A training-free, biologically-inspired method to harden AI against attacks using local noise—no need for adversarial training.