Adversarial Design Framework for Self-Driving Networks

Inspired by self-driving cars, the networking community is currently engaged in designing more automated and ``self-driving`` communication systems, aiming to overcome the cumbersome and error- prone manual approach to manage and operate networks. Ideally, such self-driving networks also allow to exploit the increasing flexibilities introduced by emerging new Internet technologies, such as software-defined and virtualized communication technologies. With these technologies, the networks allow to meet the stringent performance requirements of new applications (e.g., 5G, low-latency tele- operation, high-bandwidth machine-to-machine type communication, etc.), by adapting to the context and demand. The Internet, one of the largest and most complex artefacts built by mankind, has evolved organically over the last decades, and many design choices were taken based on experience and best practices. This project proposes a novel network framework to design and operate such networks, relying on the vision of self-driving networks, and studying how to integrate Machine Learning and Artificial Intelligence concepts into existing networks. In order to overcome the potential concerns regarding the dependability of Artificial Intelligence and Machine Learning approaches, we envision a hybrid solution which keeps the human in the loop. Hence, we first ask three fundamental questions in this project: how predictable are todays networks, i.e., user demands, workload traffic, and behavior of network functions? Can we make network design and algorithms data-driven and human interpretable? How to design a network framework that combines both generative workload models and data-driven algorithms with guarantees? The novelty of this project lies in the integration and application of Artificial Intelligence and Machine Learning on designing network algorithms. For the first time, Artificial Intelligence and Machine Learning should be integrated also in the testing and the developing phase of new networking solutions, and not only applied to solving problems. In terms of methodologies, we consider adversarial and game- theoretic approaches to test and optimize networks, to leverage the performance benefits from Machine Learning approaches while at the same time provide rigorous worst-case guarantees. Finally, a proof-of-concept implementation should demonstrate the new framework.

Communication networks have become a critical infrastructure of our digital society. As most network outages today are due to human errors, the networking community is currently engaged in designing more automated and "self-driving" communication networks that overcome today's manually managed networks. These networks exploit the flexibilities introduced by emerging software-defined communication technologies, to implement more demand-aware networks which meet the stringent requirements of new applications. The ADVISE project contributes toward our fundamental understanding of such self-driving networks, as well as first tools to realize them. To this end, we develop and apply both methods from artificial intelligence and formal approaches (and games) providing formal correctness and performance guarantees. While many of our contributions are general and of independent interest, as a case study, ADVISE focuses on emerging datacenter networks and software-defined radio access networks, two particularly critical and fast evolving types of networks. ADVISE contributions span both practical and theoretical aspects. We contribute an empirical analysis and model of the temporal and spatial structure of traffic workloads in machine-learning applications. We observe that such workloads are fairly predictable and can hence be exploited well in self-driving networks. ADVISE further contributes the algorithmic foundations for self-driving networks, leveraging and integrating predictions (as they may come from machine learning models) with formal frameworks such as competitive analysis and games. This enables novel algorithms which not only provide the classic worst-case guarantees, but which also profit from an advice which improves their performance in practice where traffic is non-adversarial, but more stochastic and predictable. For example, the ADVISE project contributes an innovative new approach, called infused advice, which allows us to analytically study and compare the performance of existing online algorithms under real workloads and prediction models, both analytically and empirically. This is very different from prior approaches in the literature, which require new algorithms that need to be tailored to a prediction model. Especially for machine-learning applications, a deep understanding of the prediction model is hard or even impossible to achieve. ADVISE also studies security aspects, identifies possible vulnerabilities of self-driving networks, and discusses how to render such networks robust. For this study, we also received a best paper award. Overall, we are very happy with the success of this project, and are deeply grateful to the FWF, also for all the support and the excellent collaboration throughout this project.