Federated Learning (FL) has become increasingly popular to perform data-driven analysis in cyber-physical critical infrastructures. Since the FL process may involve the client's confidential information, Differential Privacy (DP) has been proposed lately to secure it from adversarial inference. However, we find that while DP greatly alleviates the privacy concerns, the additional DP-noise opens a new threat for model poisoning in FL. Nonetheless, very little effort has been made in the literature to investigate this adversarial exploitation of the DP-noise. To overcome this gap, in this paper, we present a novel adaptive model poisoning technique α-MPELM} through which an attacker can exploit the additional DP-noise to evade the state-of-the-art anomaly detection techniques and prevent optimal convergence of the FL model. We evaluate our proposed attack on the state-of-the-art anomaly detection approaches in terms of detection accuracy and validation loss. The main significance of our proposed α-MPELM attack is that it reduces the state-of-the-art anomaly detection accuracy by 6.8% for norm detection, 12.6% for accuracy detection, and 13.8% for mix detection. Furthermore, we propose a Reinforcement Learning-based DP level selection process to defend α-MPELM attack. The experimental results confirm that our defense mechanism converges to an optimal privacy policy without human maneuver.