research-article

Enhancing Security-Problem-Based Deep Learning in Mobile Edge Computing

Authors:

Xiao Zheng,

Mingchu Li,

Syed Bilal Hussain Shah,

Dinh-Thuan Do,

Yuanfang Chen,

Constandinos X. Mavromoustakis,

George Mastorakis,

Evangelos PallisAuthors Info & Claims

ACM Transactions on Internet Technology (TOIT), Volume 22, Issue 2

Article No.: 49, Pages 1 - 15

https://doi.org/10.1145/3458931

Published: 14 May 2022 Publication History

Get Access

Abstract

The implementation of a variety of complex and energy-intensive mobile applications by resource-limited mobile devices (MDs) is a huge challenge. Fortunately, mobile edge computing (MEC) as a new computing paragon can offer rich resources to perform all or part of the MD’s task, which greatly reduces the energy consumption of the MD and improves the quality of service (QoS) for applications. However, offloading tasks to the edge server is vulnerable to attacks such as tampering and snooping, resulting in a deep learning (DL) security feature developed by major cloud service providers. An effective security strategy method to minimize ongoing attacks in the MEC setting is proposed. The algorithm is based on the synthetic principle of a special set of strategies, and it can quickly construct suboptimal solutions even if the number of targets achieves hundreds of millions. In addition, for a given structure and a given number of patrollers, the upper bound of the protection level can be obtained, and the lower bound required for a given protection level can also be inferred. These bounds apply to universal strategies. By comparing with the previous three basic experiments, it can be proved that our algorithm is better than the previous ones in terms of security and running time.

Appendix

Proof.

This proof is based on a comprehensive investigation of the even coverage of all targets, where the K-patrollers reflect the importance of each target. From an intuitive perspective, \(T_v\) represents the expected number of visits to v in the subsequent \(t(v)\) step. We fixed a game structure \(\Gamma\) and some patrollers \(k\in N\). Assuming that the existence of \(\eta\) follows standard parameters, then \(\eta\) is the optimal strategy for k patrollers. That is, \(Q=\prod _{v\in V} t(v), \forall 1\le i\le Q\). Assume that \(p_i(v)\) is the probability that v will occur when accessing the ith subset when the defender commits \(\eta\). For each \(1\le i \le Q\), there is \(\sum _{v\in V}P_i(v)=k\) because k patrollers are allocated in each round. Therefore, the sum of all \(P_i(v)\) is equal to \(Q\cdot k\).

\(v\in V, 0\le j\lt Q\), in which j is a multiple of \(t(v)\). Assume \(\zeta\) is the attacker’s strategy such that \(\zeta _{v,j}(w)=v\) for every history w of length j. The general calculation formula is as follows:

\begin{equation} Lev(\eta , \zeta _{v,j})=\alpha (max)-\alpha (v)\cdot \prod _{i=1}^{t(v)}(1-p\cdot P_{(i+j)}). \end{equation}

(5)

Let \(T_{v,j}=\sum _{i=1}^{t(v)}P_{i+j}(v)\). Through the analysis of Equation (5), we can see that \(Lev(\eta , \zeta _{v,j})\) is maximized if the values \(P_{i+j}(v)\) satisfy:

•

\(P_{i+j}(v)=1,\forall i=\left\lbrace 1,\ldots ,\lfloor T_{v,j} \rfloor \right\rbrace\),

•

\(P_{i+j}(v)=T_{v,j}-\lfloor T_{v,j} \rfloor , \exists i=\lfloor T_{v,j} \rfloor +1\), and

•

\(P_{i+j}(v)=0, \forall i\gt \lfloor T_{v,j} \rfloor +1\).

The right half of Equation (5) is rewritten as \(\alpha _(max)-\alpha (v)\cdot (1-p)^{\lfloor T_{v,j} \rfloor }\cdot (1-p\cdot (T_{v,j}-\lfloor T_{v,j} \rfloor))\). In the best case, for a given \(v\in V\), when all \(P_i(v)\) are optimally distributed by \(\eta\), all \(T_{v,j}\) are the same, and the same level of protection\(\eta\) is obtained for each vertex(target), so we use \(T_v\) instead of \(T_{v,j}\) to get the system \(\mathcal {L}_{\Gamma }\).

For all \(P_i(v)\)s the sum is equal to \(Q\cdot k\). By reordering these addends, you can rewrite the sum to \(\sum _{v\in V}\frac{Q}{t(v)}\cdot T_v\). So we can obtain \(k=\sum _{v\in V}\frac{T_{v}}{t(v)}\). If \(p=1\), observe that \(0\le T_{v}\le t(v)\), and we have \(0\le T_{v}\le 1\) because there are no meanings to access the same vertex multiple times.

Note here that when considering the system in solving Theorem 1, it may happen that some \(T_v\) becomes negative. If and only if the importance of certain vertices is so low that even if the discriminators do not access them at all, they are better protected than the other vertices (this explains why the sum \(\sum _{v\in V, T(v)\gt 0}\frac{T_v}{t(v)}\) is taken only from positive \(T(v)\)). The system of Theorem 1(b) may not have a reasonable solution when p is small, and the probability of each v being accessed in each step is 1, the protection \(\rho\) cannot be realized. If \(p=1\), there is no suitable solution caused by \(\rho \gt \alpha _{max}\) only.□

References

[1]

Nicola Basilico, Nicola Gatti, and Francesco Amigoni. 2009. Leader-follower strategies for robotic patrolling in environments with arbitrary topologies. In Proceedings of the 8th International Conference on Autonomous Agents and Multiagent Systems-Volume 1. International Foundation for Autonomous Agents and Multiagent Systems, 57–64.

Abstract

Appendix

References

Cited By

Index Terms

Recommendations

Modelling Task Offloading Mobile Edge Computing

Phased Scheduling for Resource-Constrained Mobile Devices in Mobile Cloud Computing

Mobile Edge Computing Services for Dynamic Quality of Service Control

Comments

Information

Published In

Publisher

Publication History

Permissions

Check for updates

Author Tags

Qualifiers

Funding Sources

Contributors

Other Metrics

Bibliometrics

Article Metrics

Other Metrics

Citations

Cited By

Login options

Full Access

View options

PDF

eReader

Full Text

HTML Format

Figures

Other

Share

Share this Publication link

Share on social media

Affiliations