Abstract: The protection of privacy against covert video recordings presents a considerable societal challenge. Our objective is to develop a computer vision system, such as a robotic device, that can identify human activities and improve daily life, all while ensuring that it does not capture video contentthat may violate individuals' privacy. In this paper, we propose a fundamental approach to reconcile these seemingly conflicting goals: the recognition of human activities using highly anonymized videodata with extremely low resolution (e.g., 16x12). We introduce a new concept called "inverse super resolution" (ISR), which entails learning the optimal set of image transformations to generate multiplelow-resolution (LR) training videos from a single highresolution source. Our ISR framework learns diverse sub-pixel transformations that are specifically optimized for activity classification, enabling the classifier to leverage high-resolution videos, such as those found on platforms like YouTube, by generating multiple LR training videos tailored to the specific activity recognition task. Through empirical experimentation, we demonstrate that the ISR paradigm significantly enhances activity recognition from extremely low-resolution video data.