Self-healing neural networks via modular patch layers for diverse structural and adversarial damages

The growing use of modern deep neural networks (DNNs) in safety-critical and continuously operating systems exposes them to safety-centric concerns and raises the usual challenge of maintaining their continued operation. Internal structural flaws or adversarial attacks may lead to performance degradation that may severely affect reliability. This paper presents a Self-Healing Neural Network (SHNN) model which can enhance recovery in post-deployment by automatically identifying, localizing, and repairing damaged components of networks with lightweight and modular patch layers. SHNN isolates faults without requiring full retraining or global fine-tuning by performing activation discrepancy analysis at each layer and retraining the afflicted layer selectively, which causes little computational and memory overhead. Experiments of MNIST, Fashion-MNIST, Sign Language MNIST, and CIFAR-10 indicate SHNN restores performance after structural damage close to baseline (97.33% and 97.22% vs. 97.76% test accuracy of the zeroed and randomised weights) and restores a lost accuracy of 74-84% in FGSM and PGD adversarial attacks. Post-healing activation deviation scores decrease significantly, confirming effective fault localization as the reduced value confirms. The framework offers a self-repair mechanism based on modules, interpretable, and dataset adaptive. with a live Streamlit interface, and allows building resilient and fault-tolerant AI systems to operate in high-reliability environments.