Developing an Original Adversarial Robustness Framework for Deep Learning Models in Nigerian Public Safety Applications

Deep learning models deployed in Nigerian public safety applications including facial recognition for border control, vehicle recognition for traffic law enforcement, and crowd behaviour analysis are susceptible to adversarial attacks that can systematically fool the model with imperceptible perturbations, yet no adversarial robustness framework has been developed for the specific threat models and operational contexts of Nigerian public safety deployments. This study developed an original adversarial robustness framework for deep learning in Nigerian public safety applications, contributing both theoretical advances and practical certification methodology. A threat model taxonomy was constructed specifically for Nigerian public safety adversarial threats, covering digital adversarial patch attacks on cameras, physical adversarial clothing and accessories, occlusion spoofing for facial recognition, and data poisoning attacks targeting model retraining pipelines. Three existing adversarial defence mechanisms (adversarial training, certified randomised smoothing, and feature squeezing) were empirically evaluated against this threat taxonomy using models trained on Nigerian datasets for each application domain. Adversarial training showed the best empirical robustness under physical patch attacks (92.3 percent maintained accuracy versus 41.7 percent without defence), while randomised smoothing provided the only certifiable defence guarantee (provably correct predictions within an L2 perturbation radius of 0.5 for 78.4 percent of test examples). An original Adaptive Robustness Certification Scheme was developed, combining empirical worst-case testing with certifiable smooth classifiers to provide a practical assurance level for public safety deployment decisions. Expert review by 18 adversarial ML and public safety system specialists confirmed the framework's original contribution.

Keywords: adversarial robustness, deep learning, public safety, Nigeria, adversarial attacks