Adversarial Explanations for Informed Civilian and Environmental Protection

Combating crime and conditions of high physical risk in cities, the environment, and critical infrastructures requires a multifaceted approach. For sensitive problems, such as advanced situational awareness in the fields of civilian applications and environmental protection, Artificial Intelligence (AI) and Neural Network (NN) adoption has been slow due to concerns about their reliability, leading to several algorithms for explaining their decisions. Despite the possibilities for AI in critical infrastructure protection and civilian applications, many challenges still exist. For instance: (i) there are complex and high risks meaning that AI systems need to be transparent and interpretable to gain decision-maker trust; (ii) AI models may be vulnerable to imperceptible manipulations of input data even without any knowledge about the AI technique that is used; (iii) the need to efficiently process distributed, multimodal and big data coming from different, but however cheap, Internet of Things (IoT) and sensory devices (e.g., drones, cameras, accelerometers, telemetry, geomagnetic field, and proximity sensors); and (iv) many AI methods based on Machine Learning (ML) require huge amounts of training data, resulting in a Big Data computation problem. We introduce, benchmark, and demonstrate an adversarial explanations approach that we can efficiently tackle both adversarial robustness and explanation complexity of AI systems. To achieve this, we train robustified NNs and transparent explainers on big imagery data and leverage the attacks’ knowledge as explanations to gain greater fidelity to the AI model. The merit of the proposed approach is that the new and robustified model has a great performance against new, unseen types of perturbations and attacks. This way, we pave the adoption of more informed and responsible AI integration in sensitive application domains.