Raluca Ada Popa — Page 3

arXiv2017

A Berkeley View of Systems Challenges for AI

With the increasing commoditization of computer vision, speech recognition and machine translation systems and the widespread deployment of learning-based back-end technologies such as digital advertising and intelligent infrastructures, AI (Artificial Intelligence) has moved from research labs to production. These changes have been made possible by unprecedented levels of data and computation, by methodological advances in machine learning, by innovations in systems software and architectures, and by the broad accessibility of these technologies. The next generation of AI systems promises to accelerate these developments and increasingly impact our lives via frequent interactions and making (often mission-critical) decisions on our behalf, often in highly personalized contexts. Realizing this promise, however, raises daunting challenges. In particular, we need AI systems that make timely and safe decisions in unpredictable environments, that are robust against sophisticated adversaries, and that can process ever increasing amounts of data across organizations and individuals without compromising confidentiality. These challenges will be exacerbated by the end of the Moore's Law, which will constrain the amount of data these technologies can store and process. In this paper, we propose several open research directions in systems, architectures, and security that can address these challenges and help unlock AI's potential to improve lives and society.

cs.AI↗

arXiv2011

Going Beyond Pollution Attacks: Forcing Byzantine Clients to Code Correctly

Network coding achieves optimal throughput in multicast networks. However, throughput optimality \emph{relies} on the network nodes or routers to code \emph{correctly}. A Byzantine node may introduce junk packets in the network (thus polluting downstream packets and causing the sinks to receive the wrong data) or may choose coding coefficients in a way that significantly reduces the throughput of the network. Most prior work focused on the problem of Byzantine nodes polluting packets. However, even if a Byzantine node does not pollute packets, he can still affect significantly the throughput of the network by not coding correctly. No previous work attempted to verify if a certain node \emph{coded correctly using random coefficients} over \emph{all} of the packets he was supposed to code over. We provide two novel protocols (which we call PIP and Log-PIP) for detecting whether a node coded correctly over all the packets received (i.e., according to a random linear network coding algorithm). Our protocols enable any node in the network to examine a packet received from another node by running a "verification test". With our protocols, the worst an adversary can do and still pass the packet verification test is in fact equivalent to random linear network coding, which has been shown to be optimal in multicast networks. Our protocols resist collusion among nodes and are applicable to a variety of settings. Our topology simulations show that the throughput in the worst case for our protocol is two to three times larger than the throughput in various adversarial strategies allowed by prior work. We implemented our protocols in C/C++ and Java, as well as incorporated them on the Android platform (Nexus One). Our evaluation shows that our protocols impose modest overhead.

cs.NI↗