Neural Network Topology Reconfiguration Against Backdoor Attacks

The patent extracts and removes only the dormant neurons and links that are susceptible to cyber attacks, rather than removing random or excessive connections. This selective extraction approach eliminates vulnerable components while preserving the functional integrity and learning accuracy of the neural network.

The patent identifies and removes vulnerable neurons and links before the neural network is deployed or before cyber attacks occur. This preliminary action prevents potential attack vectors from being established, thereby improving robustness without requiring post-attack repairs or accuracy compromises.

The patent implements a scale-free topology where neurons have asymmetric connection degrees following a power-law distribution. Some neurons have many connections (hubs) while others have few, creating an asymmetric structure that is robust against random failures and targeted attacks on individual nodes, while avoiding uniform complexity across all connections.

The patent changes the connectivity parameter distribution from random or regular patterns to a scale-free power-law distribution. This parameter change optimizes the balance between robustness and complexity by concentrating connections on critical neurons while maintaining sparsity elsewhere, thereby improving attack resistance without proportionally increasing overall complexity.