Page 27 - ITU Journal Future and evolving technologies – Volume 2 (2021), Issue 2
P. 27
ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 2
task based on the load ratio of each provider, and apply
task scheduling, which initializes main parameters, and reinforcement learning techniques to protect edges
uses the precedent parts of the algorithm to ind the from ing, malware, jamming, and eavesdropping
best ob‑ jective function for each task. The main issue attacks that might occur during data of loading to edges
associated with this solution is the fact that it is a nodes. The radiocommunication channels of edges nodes
static algorithm and it is suitable for batch scheduling are vulnerable to attacks launched from the physical
only. layer or Medium Access Control (MAC) layers during data
loading in MCC environment. Most of the solutions
Li et al. [50] proposed a computation include the use of Q‑learning to prevent attacks; the main
partitioning technique to improve the performance of reason is the fact that Q‑learning‑based security schemes
big data en‑ vironments in MCC. In this technique, do not require any prior knowledge of the network, they
two computation partitions are considered; the irst apply the iterative Bellman equation to update Q‑values,
one is responsible for monitoring the changes in and only use two parameters, which are the learning
resources including bandwidth, CPU, etc., while the rate and the discount factor, to control their learning
other partition deter‑ mines the computation performance. Nonetheless, security schemes based on
location where a given task should be executed. Q‑learning require exploring all the possible states and
The main goal of this methodology is to enhance the pairs of actions before signi icantly changing the network
application performance on UE by improving the policies, resulting in a slower reaction in case of an
computation partitioning decision. For this, it is imminent attack.
important to ind an ef icient way to solve the
single‑frame execution time problem, then Nguyen et al. [53] proposed a method based on deep
establish the partitioning scheme for multi‑frame learning to prevent and detect cyberattacks in MCC: a
execution, as the single‑frame execution alone is training dataset is used to train the neural networks of
icient. These calculations depend on the network the framework that implements the technique of line,
bandwidth and the changes in the environment of the then, once the model is ready, it is integrated in the
system. The model includes three types of tasks that MCC environment to detect and prevent attacks online.
are local, transferring and cloud tasks. A graph is used The model involves two major phases which are feature
to represent the tasks’ data low, and the adjacency analysis and learning process. Feature analysis includes
matrix of the graph is used to perform task selection. the extraction and examination of abnormal attributes in
For the single‑frame task execution problem, the the dataset to identify traits associated with malicious
icient partitioning scheme is determined by a packets, and dimensionality reduction using the Prin‑
Genetic Algorithm (GA) due to its strong search cipal Component Analysis (PCA) technique to remove
capabilities. Additional optimization and adjustments irrelevant features or attributes that are not needed for
are performed to settle the total execution time of the detection of attacks. The learning process comprises
multi‑frame data. However, for this solution to be three types of layers including the input layer, some
effective, data‑frames congestion, instability of data hidden layers and the output layer. The features are
during transfer, and limitation of resources should fed directly to the input layer; then, a Gaussian Binary
be considered. Restricted Boltzmann Machine (GRBM) is used to convert
them into binary codes, which are used in the hidden
4.5 Solutions for data security and privacy layers. A series of learning steps are performed to adjust
the weights of each layer. However, only theoretical
Qiu, Tie, et al. [51] proposed SIGMM, a machine evaluation of the model was performed, even though high
learning algorithm for spammer identi ication in accuracy was obtained, the model was not evaluated in a
industrial MCC. The framework makes use of data, practical and real time environment.
where each user node is ied into one class in
the construction process of the model, the data To improve the iciency of encryption and decryp‑
includes the relationship with other users, user’s tion schemes in MCC and make them suitable for mobile
identi ication, the time‑stamped post record, and the devices, Zhang et al. [52] introduced a system archi‑
activity in the past three months. A Pearson tecture of anonymous attribute‑based access control in
correlation coef icient and Principal Component mobile cloud computing, a decryption method called
Analysis (PCA) were employed to characterize match‑then‑decrypt where a matching phase is added
different features and model the parameters before the decryption phase. The technique involves a
accurately. SIGMM its the behavior data of regular basic anonymous Cyphertext Policy ‑ Attribute‑Based
users and spammers, in which the behavior data of Encryption (CP‑ABE) construction and the procurement
ordinary users and spammers are mixed by random of security‑enhanced extension using the reasonable
sampling. However, this solution is not suitable for Canetti–Halevi–Katz technique based on one‑time signa‑
large networks since the algorithm is based on tures. In Canetti –Halevi –Katz transformation, a test can
binary ication, the types of users are varied and be made during the decryption process before complet‑
complex in large networks and thus more than two ing it and the subsequent decryption is completed if and
categories are required to classify the nodes
accurately.
© International Telecommunication Union, 2021 13