Neurodynamics-inspired dynamic malware detection framework with synergistic small-sample adaptation and robust defense

This study proposes a neuro-dynamics–inspired malicious app detection framework designed to address the dual challenges of small-sample learning and adversarial attacks. By introducing synaptic plasticity–driven dynamic meta-learning, the framework improves detection accuracy by 28% points, reaching 93.7% in 5-way 1-shot tasks on the Drebin dataset, compared to 65.2% for conventional graph neural networks. A Gamma-band neural oscillation–based adversarial immunity mechanism is incorporated to enhance robustness: it reduces the PGD attack success rate from 98% to 12%, expands the certified authentication radius by 2.3×, and maintains an average accuracy of 78.4% under six types of adversarial attacks (FGSM, PGD, C&W, BIM, DeepFool, AE-Sign). To enable efficient deployment, a dynamic pruning + spiking neuron encoding strategy compresses model parameters by 63%, achieving real-time detection at 28 ms with only 6.5 W power consumption on ARM + NPU architecture. In cross-domain evaluations, the framework achieves a zero-day attack detection rate of 96.2% in smart city scenarios, and maintains a false positive rate below 0.5% in healthcare environments—both significantly outperforming traditional baselines. These results demonstrate that the proposed framework simultaneously improves detection accuracy, adversarial robustness, and edge deployment efficiency, shifting malicious app detection from a static rule-driven paradigm to a dynamic, neuro-inspired security paradigm suitable for high-risk intelligent terminal environments. 28% Boost: hits 93.7% on Drebin in 5-way 1-shot few-shot learning. PGD Shield: slashes attack success from 98% → 12% and widens certified radius 2.3×. Edge-Ready: 28 ms latency, 6.5 W power—dynamic pruning + spike coding for real-time IoT defense.