Editing Emerging Research Directions (section)

===7.5.6 Fault Tolerance and Recovery===

'''Byzantine Faults at the Edge'''

Byzantine faults refer to an issue where in some distributed system, nodes (devices) may behave un predictably due to failures or attacks. Faulty nodes may send incorrect data which results in data corruption that affects other nodes in the server. This problem is essential to explore in edge systems as one main challenge of edge computing is the unreliability of resource constraint devices. Devices located to data sources tend to be highly unpredictable (e.g. IoT devices, mobile phones, or embedded systems). There are many algorithmic solutions to reduce Byzantine faults such as Bracha-Toueg Byzantine consensus algorithm, Mahaney-Schneider synchronizer, and other consensus algorithms, 

[[File:Bracha-Toueg.png|600px|thumb|center| Example of consensus algorithm, Bracha-Toueg Byzantine consensus algorithm.]]

''Byzantine Generals Problem, which was introduced by Leslie Lamport, Robert Shostak, and Marshall Pease in their 1982 paper titled The Byzantine Generals Problem''

'''Case Study 1: Byzantine resilience in federate learning at the edge'''

Creative solutions in edge systems are explored in [1], Byzantine resilience in federated learning at the edge, where heavy tailed data and communication inefficiencies are main challenges presented in this paper. Federated learning is a prominent example of a edge tailored system where Byzantine resilience is essential. Authors of this paper proposed a distributed gradient descent algorithm that operates with local computations on each device where it calculates the gradient of the model based on local data. In order to trim failed nodes, the algorithm proposed uses ''coordinate-wise trimmed mean''. This algorithm is designed to follow heavy tailed distribution which ensures outliers do not affect the model. The key issues addressed in the paper is how Byzantine failures degrade the performance of federated learning systems. This solution employes robust aggregation in order to solve these issues where traditional algorithms are inapplicable.