Federated Learning (FL) aims to improve machine learning privacy by allowing several data owners in edge and ubiquitous computing systems to collaboratively train a model, while preserving their local training data private, and sharing only model training parameters. However, FL systems remain vulnerable to privacy attacks, and in particular, to membership inference attacks that allow adversaries to determine whether a given data sample belongs to participants' training data, thus, raising a significant threat in sensitive ubiquitous computing systems. Indeed, membership inference attacks are based on a binary classifier that is able to differentiate between member data samples used to train a model and non-member data samples not used for training. In this context, several defense mechanisms, including differential privacy, have been proposed to counter such privacy attacks. However, the main drawback of these methods is that they may reduce model accuracy while incurring non-negligible computational costs. In this paper, we precisely address this problem with PASTEL, a FL privacy-preserving mechanism that is based on a novel multi-objective learning function. On the one hand, PASTEL decreases the generalization gap to reduce the difference between member data and non-member data, and on the other hand, PASTEL reduces model loss and leverages adaptive gradient descent optimization for preserving high model accuracy. Our experimental evaluations conducted on eight widely used datasets and five model architectures show that PASTEL significantly reduces membership inference attack success rates by up to -28%, reaching optimal privacy protection in most cases, with low to no perceptible impact on model accuracy.