We want to study a new adversarial deep learning framework for the problem of human activity recognition (HAR) on people wearing inertial sensors. Our framework aims to study a new formulation of an adversarial task based on the concept of inter-person variability, i.e., different people perform the same activity in different ways. We aim to explore how this new discrimination task improves the classification performance in comparison to discrimination tasks in previous works while reducing the variability per person and activity. Overall, we will explore the behaviour of our framework in several Human Activity Recognition public datasets using a leave-one(person)-out cross-validation strategy.