Safe-DL: A Modular Framework for Evaluating and Mitigating Adversarial Threats in Deep Learning

Abstract

The increasing adoption of deep neural network-based systems in critical domains, such as medicine, mobility, and security, raises serious concerns regarding their robustness against adversarial threats. Despite their strong performance, these systems exhibit exploitable vulnerabilities that may compromise automated decision-making processes. While partial solutions to mitigate such threats exist, the current ecosystem is characterised by fragmented, complex, and hard-to-integrate approaches, often disconnected from emerging regulatory guidelines such as the European AI Act. This dissertation introduces Safe-DL, a modular, extensible, and auditable framework designed to evaluate and mitigate adversarial threats in deep learning systems. The solution is organised into six modules: threat modelling, attack simulation, risk analysis, defence application, mitigation assessment, and final reporting. Together, these modules enable a systematic and integrated security analysis of machine learning models. Experimental validation was carried out in a realistic image classification scenario, including data poisoning, backdoor, and evasion attacks. The results demonstrate the framework's ability to identify vulnerabilities and apply effective countermeasures, producing traceable and human-readable outputs. By combining reproducibility, usability, and alignment with ethical and legal requirements, this framework offers a practical and scientific contribution toward building more robust, transparent, and trustworthy AI systems. The work developed using Safe-DL lays a solid foundation for future extensions of the framework to other modalities and regulatory contexts.