Securing Deep Learning Systems: Attacks and Defenses across Neural Networks, Federated, Transfer, and Reinforcement Learning

Abstract

Deep Learning (DL) techniques are widely deployed in critical applications such as autonomous driving, healthcare, and intelligent infrastructure. Core paradigms—Deep Neural Networks (DNNs), Deep Reinforcement Learning (DRL), Federated Learning (FL), and Transfer Learning (TL)—remain vulnerable to adversarial attacks that can degrade performance, leak private data, or produce unsafe decisions. Developing effective attacks and corresponding countermeasures is a prerequisite for robust, secure, and deployable artificial intelligence. Prior surveys often focused on only one or two techniques, omitted detailed discussion of datasets, metrics, and testbeds, or became outdated. This survey comprehensively reviews attacks and defenses across DNN, DRL, FL, and TL. We summarize threat models, representative attack and defense techniques, evaluation metrics, commonly used datasets, and experimental settings. A key contribution is an explicit analysis of the commonalities and differences among the four paradigms. Insights, lessons learned, and future research directions are presented to guide the development of trustworthy deep-learning systems.