Classification-based prediction of network connectivity robustness.

Area of Science:

• Network Science
• Computer Science
• Artificial Intelligence

Background:

• Network robustness is critical for industrial and societal security against malicious attacks.
• Evaluating network connectivity robustness traditionally involves time-consuming attack simulations, especially for large-scale systems.
• Existing methods for measuring network robustness are computationally intensive.

Purpose of the Study:

• To propose an efficient predictor for network connectivity robustness using multiple convolutional neural networks (mCNN-RP).
• To extend single CNN-based predictors for enhanced robustness estimation.
• To reduce the computational burden of assessing network connectivity robustness.

Main Methods:

• Developed a multiple convolutional neural network predictor (mCNN-RP) for network connectivity robustness.
• Utilized one CNN as a classifier and other CNNs as estimators for classified network categories.
• Incorporated a data-based filter for refining predictive data.
• Conducted experiments on synthetic and real-world networks (directed, undirected, weighted, unweighted).

Main Results:

• mCNN-RP demonstrated effectiveness across diverse network topologies.
• Achieved an average prediction error lower than the standard deviation of the data.
• Significantly reduced the runtime for assessing network connectivity robustness.
• Outperformed single CNN-based frameworks and other existing prediction measures.

Conclusions:

• mCNN-RP accurately predicts connectivity robustness in complex networks.
• The proposed method offers a superior indicator for network connectivity robustness.
• mCNN-RP provides a computationally efficient alternative to traditional robustness assessment techniques.