Chaos-driven hyperchaotic synchronization framework for enhancing neural network security

Deep neural networks (DNNs) remain highly vulnerable to adversarial attacks, where imperceptible perturbations significantly degrade classification accuracy. Existing defense strategies, including adversarial training and defensive distillation, often fail under adaptive and strong adversarial settings. This paper proposes a novel Chaos-Driven Neural Network Security (CNNS) framework that leverages a four-dimensional (4D) hyperchaotic system as a dynamic perturbation generator, coupled with an adaptive sliding-mode master-slave synchronization controller that guarantees robust, bounded, and reproducible chaotic signals under parametric uncertainties. The hyperchaotic signal is injected simultaneously into input data, hidden activations, and network weights via a dimensionality-consistent projection operator, inducing non-stationary decision boundaries that disrupt gradient-based attack optimization. A rigorous Lyapunov stability proof, including the explicit adaptive law for uncertainty estimation, guarantees asymptotic convergence of the synchronization error to zero. Robustness is validated on MNIST and CIFAR-10 against standard attacks (FGSM, PGD, DeepFool, CW) as well as adaptive attacks (BPDA+PGD), achieving 79.2% and 81.4% accuracy under FGSM (ε = 0.05), respectively, outperforming state-of-the-art defense methods.