Résumé

Artificial Intelligence (AI) is revolutionizing information security, influencing both attack and defense strategies. Attackers leverage AI to automate cyberattacks and exploit vulnerabilities, while defenders utilize it for anomaly detection, predictive threat modeling, and automated responses. Federated Learning (FL), a privacy-preserving training method, remains vulnerable to inference attacks. To address this, we propose the Rényi Differential Privacy (RDP) based federated learning (RDP-FL) framework, which incorporates moment accounted noise scaling to dynamically regulate the privacy budget, achieving an optimal balance between privacy and utility. This method minimizes unnecessary noise addition while maintaining strong privacy guarantees, thereby preserving data integrity and enhancing model performance. Experimental validation on the Medical-MNIST and CIFAR-10 datasets demonstrates the effectiveness of RDP-FL, showing its ability to safeguard data privacy while ensuring high classification accuracy. This work advances the ongoing efforts to enhance cybersecurity in an AI-driven landscape.