Aika: A Distributed Edge System For Machine Learning Inference. Detecting and defending against abnormal behavior in untrusted edge environments

Abstract

The edge computing paradigm has recently started to gain a lot of momentum. The field of Artificial Intelligence (AI) has also grown in recent years, and there is currently ongoing research that investigates how AI can be applied to numerous of different fields. This includes the edge computing domain. In Norway, there is currently ongoing research being conducted that investigates how the confluence between AI and edge computing can be used to hinder fish crime, by stationing surveillance equipment aboard fishing vessels, and perform all the monitoring directly on the vessel with support of AI.

This is challenging for several reasons. First and foremost, the equipment needs to be stationed on the vessel, where actors may impose a threat to it and attempt to damage it, or interfere with the analytical process. The second challenge is to enable multiple machine learning pipelines to be executed effectively on the equipment. This requires a versatile computation model, where data is handled in a privacy preserving manner.

This thesis presents Áika, a distributed edge computing system that supports machine learning inference in such untrusted edge environments. Áika is designed as a hierarchical fault tolerant system that supports a directed acyclic graph (DAG) computation model for executing machine inference on the edge, where a monitor residing in a trusted location can ensure that the system is running as expected.

The experiment results demonstrate that Áika can tolerate failures while remaining operable with a stable throughput, although this will depend on the specific configuration and what computations that are implemented. The results also demonstrate that Áika can be used for both simple tasks, like counting words in a textual document, and for more complex tasks, like performing feature extraction using pre-trained deep learning models that are distributed across different workers. With Áika, application developers can develop fault tolerant and different distributed DAGs composed of multiple pipelines.