Activity: Talk or presentation types › Invited talk
Description
Due to the recent breakthroughs in Machine Learning (ML), ML-powered intelligent systems are being widely deployed in mainstream as well as industrial application domains, and applied to solving a range of complex real-life problems. Two of the primary challenges to the deployment of such systems, especially in the context of embedded systems and the Edge are: (1) Security, and (2) Power consumption and resource utilization. In fact, these systems suffer from persistent vulnerability to adversarial attacks, i.e., carefully crafted additive noise that forces the model to output wrong labels. Moreover, state-of-the-art defenses against these attacks result in high overheads in terms of energy, time and resources. This talk will give an overview on real-world adversarial attacks as well as a new family of defenses that leverages Approximate Computing paradigm to jointly tackle these challenges.