1.
Manisha Milind Patil
– Research Scholar, Balaji Institute of Modern Management, Sri Balaji University, Pune, Maharashtra, India.
2.
Dhanraj P. Tambuskar
– Professor, Balaji Institute of Modern Management, Sri Balaji University, Pune, Maharashtra, India.
Abstract
Cloud computing has emerged as the primary model for providing
flexible computing resources, with applications ranging from
common enterprise systems to those that are safety critical.
However, the very aspects of cloud computing such as scalability
and elasticity complicate the monitoring process, since jobs are
transferred from one server to another, resource baselines change
due to automatic resource scaling, and the amount of data to
be analysed increases significantly. In this connection, anomaly
detection technology, comprising techniques that help identify
unusual activities in the system, has been developed and applied
within the framework of cloud computing. This article conducts
a systematic review of the existing literature, in line with the
recommendations provided by Kitchenham. The review aims to
answer four questions relating to the types of anomalies studied,
the techniques chosen for analysis, the types of data used, and
the factors that influence the choice of a particular methodology.
Following a systematic search of five digital libraries (IEEE Xplore,
Scopus, ACM Digital Library, ScienceDirect, and SpringerLink)
from 2015 to 2026, duplicates were removed, and studies were
subjected to two-stage screening and quality assessment to yield
78 primary studies for the subsequent qualitative synthesis. In
this way, we establish a taxonomy of cloud anomalies in three
dimensions, namely performance, security, and resources,
and conduct a comparison of four major categories of anomaly
detection techniques, i.e., statistical approaches, classical
machine learning, unsupervised learning, and deep learning, as
well as hybrids and ensembles, based on criteria that matter in
a production cloud environment, i.e., detection accuracy, falsepositive
rate, computational cost, adaptability to concept drift,
interpretability, and streaming capability. It is evident from our
findings that there is no one-size-fits-all solution, since although deep learning algorithms like LSTM autoencoders provide high
accuracy in detecting anomalies in highly complex temporal
telemetry data, these solutions are highly costly and non-intuitive,
whereas light-weight algorithms like Isolation Forest and Statistical
Process Control manage to deliver satisfactory performance in
real-time and constrained environments.
Keywords Anomaly Detection, Cloud Computing, Systematic Literature Review, Machine Learning, Deep Learning, Intrusion Detection, Performance Monitoring, AIOps