Boosted training strategy for physics-informed neural networks in modeling non-linear computer virus dynamics with vertical transmission

Abstract

Computer-virus propagation in networked systems is often modeled by epidemic-type derivative equations. Still, standard quantitative solvers can become computationally demanding and may exhibit small stability for long-term predictions. This work develops a Physics-Informed Neural Network (PINN) based framework for an e-epidemic SI1I2R model of computer-virus spread that incorporates a possibly transmissible class, an amply transmissible class, and direct transmission, allowing nodes to be initially compromised without contact. The proposed PINN uses a multilayer, amply connected architecture with Swish-ReLU activations to approximate the solution to the coupled nonlinear common differential equations and to enforce the governing dynamics through a physics-based loss. We use standard error metrics such as Mean Squared Error (MSE), Theil’s Inequality Coefficient (TIC), and Mean Absolute Deviation (MAD) to compare the model's performance against a reference ODE Solver. We also examine training and validation loss profiles across several trials. A grid-based sensitivity analysis of key parameters (the infection rate and the progression rate from I1 to I2) is conducted to determine how they affect the peak sizes of infected and recovered populations. The results show that the PINN accurately reproduces the reference SI1I2R dynamics with low errors across all compartments. Swish-ReLU is the most stable and accurate training function of the ones that were tested. These results indicate that PINNs are a reliable and possibly resource-efficient way to model and predict how computer viruses spread. They also show that they could be useful for future AI-based epidemic modeling in cybersecurity.