Explanation-Driven Adversarial Attacks against Multimedia Edge Applications

Generative Artificial Intelligence (GenAI) and automated attacks on AI models raise concerns about the credibility and reliability of AI systems. Specifically, for civilian applications, there is an increasing need for AI systems to be robust, transparent, and interpretable to earn the trust of decision-makers. To effectively address the challenge of learning with enhanced explanations in constrained edge applications under conditions of massive adversarial attacks, we benchmark and present an Explanation-Driven Adversarial approach. By harvesting multi-media data collected from drones at the edge and augmenting it with diverse and massive adversarial examples, we transparently train explainers and robustify Neural Networks (NNs) to improve the AI model’s fidelity. The new and robust AI model performs exceptionally well against novel and unseen attack types and concept drifts. Finally, through benchmarking across diverse adversarial attacks, we extend research in sensitive application domains and promote the adoption of more responsible and informed AI integration.