Self-Repairing Neural Networks based on Neural Cellular Automata

Abstract

We propose using the Growing Neural Cellular Automata (GNCA) framework to repair neural networks subjected to weight tampering and removal. Each weight of a dense layer is treated as a cell in a two-dimensional grid, and a neural CA is trained to regenerate the weight matrix from localised corruption using only local cell interactions — with no centralised copy of the parameters. Motivated by cybersecurity applications such as adversarial robustness and model extraction defence, we evaluate the approach on a TinyMLP trained on MNIST. The NCA retains 88.8% classification accuracy at 90% weight corruption, while standard baselines approach random guessing. We outline connections to fault-tolerant computation and directions for extension to multi-layer networks and adversarial damage models.