Loading [MathJax]/jax/element/mml/optable/GeneralPunctuation.js
Research article Special Issues

Personalized federated learning for heterogeneous data: A distributed edge clustering approach


  • Federated learning (FL) is a distributed machine learning technique that allows multiple devices (e.g., smartphones and IoT devices) to collaborate in the training of a shared model with each device preserving the privacy of its local data. However, the highly heterogeneous distribution of data among clients in FL can result in poor convergence. In addressing this issue, the concept of personalized federated learning (PFL) has emerged. PFL aims to tackle the effects of non-independent and identically distributed data and statistical heterogeneity and to achieve personalized models with rapid model convergence. One approach is clustering-based PFL, which utilizes group-level client relationships to achieve personalization. However, this method still relies on a centralized approach, whereby the server coordinates all processes. To address these shortcomings, this study introduces a blockchain-enabled distributed edge cluster for PFL (BPFL) that combines the benefits of blockchain and edge computing. Blockchain technology can be used to enhance client privacy and security by recording transactions on immutable distributed ledger networks, thereby improving client selection and clustering. The edge computing system offers reliable storage and computation such that computational processing is locally performed in the edge infrastructure to be closer to clients. Thus, the real-time services and low-latency communication of PFL are improved. However, further work is required to develop a representative dataset for the examination of related types of attacks and defenses for a robust BPFL protocol.

    Citation: Muhammad Firdaus, Siwan Noh, Zhuohao Qian, Harashta Tatimma Larasati, Kyung-Hyune Rhee. Personalized federated learning for heterogeneous data: A distributed edge clustering approach[J]. Mathematical Biosciences and Engineering, 2023, 20(6): 10725-10740. doi: 10.3934/mbe.2023475

    Related Papers:

    [1] Tianyuan Xu, Shanming Ji, Chunhua Jin, Ming Mei, Jingxue Yin . EARLY AND LATE STAGE PROFILES FOR A CHEMOTAXIS MODEL WITH DENSITY-DEPENDENT JUMP PROBABILITY. Mathematical Biosciences and Engineering, 2018, 15(6): 1345-1385. doi: 10.3934/mbe.2018062
    [2] Wenjie Zhang, Lu Xu, Qiao Xin . Global boundedness of a higher-dimensional chemotaxis system on alopecia areata. Mathematical Biosciences and Engineering, 2023, 20(5): 7922-7942. doi: 10.3934/mbe.2023343
    [3] Sunwoo Hwang, Seongwon Lee, Hyung Ju Hwang . Neural network approach to data-driven estimation of chemotactic sensitivity in the Keller-Segel model. Mathematical Biosciences and Engineering, 2021, 18(6): 8524-8534. doi: 10.3934/mbe.2021421
    [4] Qianhong Zhang, Fubiao Lin, Xiaoying Zhong . On discrete time Beverton-Holt population model with fuzzy environment. Mathematical Biosciences and Engineering, 2019, 16(3): 1471-1488. doi: 10.3934/mbe.2019071
    [5] Chichia Chiu, Jui-Ling Yu . An optimal adaptive time-stepping scheme for solving reaction-diffusion-chemotaxis systems. Mathematical Biosciences and Engineering, 2007, 4(2): 187-203. doi: 10.3934/mbe.2007.4.187
    [6] Xu Song, Jingyu Li . Asymptotic stability of spiky steady states for a singular chemotaxis model with signal-suppressed motility. Mathematical Biosciences and Engineering, 2022, 19(12): 13988-14028. doi: 10.3934/mbe.2022652
    [7] Tingting Yu, Sanling Yuan . Dynamics of a stochastic turbidostat model with sampled and delayed measurements. Mathematical Biosciences and Engineering, 2023, 20(4): 6215-6236. doi: 10.3934/mbe.2023268
    [8] Lin Zhang, Yongbin Ge, Zhi Wang . Positivity-preserving high-order compact difference method for the Keller-Segel chemotaxis model. Mathematical Biosciences and Engineering, 2022, 19(7): 6764-6794. doi: 10.3934/mbe.2022319
    [9] Changwook Yoon, Sewoong Kim, Hyung Ju Hwang . Global well-posedness and pattern formations of the immune system induced by chemotaxis. Mathematical Biosciences and Engineering, 2020, 17(4): 3426-3449. doi: 10.3934/mbe.2020194
    [10] Marcin Choiński, Mariusz Bodzioch, Urszula Foryś . A non-standard discretized SIS model of epidemics. Mathematical Biosciences and Engineering, 2022, 19(1): 115-133. doi: 10.3934/mbe.2022006
  • Federated learning (FL) is a distributed machine learning technique that allows multiple devices (e.g., smartphones and IoT devices) to collaborate in the training of a shared model with each device preserving the privacy of its local data. However, the highly heterogeneous distribution of data among clients in FL can result in poor convergence. In addressing this issue, the concept of personalized federated learning (PFL) has emerged. PFL aims to tackle the effects of non-independent and identically distributed data and statistical heterogeneity and to achieve personalized models with rapid model convergence. One approach is clustering-based PFL, which utilizes group-level client relationships to achieve personalization. However, this method still relies on a centralized approach, whereby the server coordinates all processes. To address these shortcomings, this study introduces a blockchain-enabled distributed edge cluster for PFL (BPFL) that combines the benefits of blockchain and edge computing. Blockchain technology can be used to enhance client privacy and security by recording transactions on immutable distributed ledger networks, thereby improving client selection and clustering. The edge computing system offers reliable storage and computation such that computational processing is locally performed in the edge infrastructure to be closer to clients. Thus, the real-time services and low-latency communication of PFL are improved. However, further work is required to develop a representative dataset for the examination of related types of attacks and defenses for a robust BPFL protocol.



    Baló's concentric sclerosis (BCS) was first described by Marburg [1] in 1906, but became more widely known until 1928 when the Hungarian neuropathologist Josef Baló published a report of a 23-year-old student with right hemiparesis, aphasia, and papilledema, who at autopsy had several lesions of the cerebral white matter, with an unusual concentric pattern of demyelination [2]. Traditionally, BCS is often regarded as a rare variant of multiple sclerosis (MS). Clinically, BCS is most often characterized by an acute onset with steady progression to major disability and death with months, thus resembling Marburg's acute MS [3,4]. Its pathological hallmarks are oligodendrocyte loss and large demyelinated lesions characterized by the annual ring-like alternating pattern of demyelinating and myelin-preserved regions. In [5], the authors found that tissue preconditioning might explain why Baló lesions develop a concentric pattern. According to the tissue preconditioning theory and the analogies between Baló's sclerosis and the Liesegang periodic precipitation phenomenon, Khonsari and Calvez [6] established the following chemotaxis model

    ˜uτ=DΔX˜udiffusion ofactivated macrophagesX(˜χ˜u(ˉu˜u)˜v)chemoattractant attractssurrounding activated macrophages+μ˜u(ˉu˜u)production of activated macrophages,˜ϵΔX˜vdiffusion of chemoattractant=˜α˜v+˜β˜wdegradationproduction of chemoattractant,˜wτ=κ˜uˉu+˜u˜u(ˉw˜w)destruction of oligodendrocytes, (1.1)

    where ˜u, ˜v and ˜w are, respectively, the density of activated macrophages, the concentration of chemoattractants and density of destroyed oligodendrocytes. ˉu and ˉw represent the characteristic densities of macrophages and oligodendrocytes respectively.

    By numerical simulation, the authors in [6,7] indicated that model (1.1) only produces heterogeneous concentric demyelination and homogeneous demyelinated plaques as χ value gradually increases. In addition to the chemoattractant produced by destroyed oligodendrocytes, "classically activated'' M1 microglia also can release cytotoxicity [8]. Therefore we introduce a linear production term into the second equation of model (1.1), and establish the following BCS chemotaxis model with linear production term

    {˜uτ=DΔX˜uX(˜χ˜u(ˉu˜u)˜v)+μ˜u(ˉu˜u),˜ϵΔX˜v+˜α˜v=˜β˜w+˜γ˜u,˜wτ=κ˜uˉu+˜u˜u(ˉw˜w). (1.2)

    Before going to details, let us simplify model (1.2) with the following scaling

    u=˜uˉu,v=μˉu˜ϵD˜v,w=˜wˉw,t=μˉuτ,x=μˉuDX,χ=˜χ˜ϵμ,α=D˜α˜ϵμˉu,β=˜βˉw,γ=˜γˉu,δ=κμ,

    then model (1.2) takes the form

    {ut=Δu(χu(1u)v)+u(1u),xΩ,t>0,Δv+αv=βw+γu,xΩ,t>0,wt=δu1+uu(1w),xΩ,t>0,ηu=ηv=0,xΩ,t>0,u(x,0)=u0(x),w(x,0)=w0(x),xΩ, (1.3)

    where ΩRn(n1) is a smooth bounded domain, η is the outward normal vector to Ω, η=/η, δ balances the speed of the front and the intensity of the macrophages in damaging the myelin. The parameters χ,α and δ are positive constants as well as β,γ are nonnegative constants.

    If δ=0, then model (1.3) is a parabolic-elliptic chemotaxis system with volume-filling effect and logistic source. In order to be more line with biologically realistic mechanisms, Hillen and Painter [9,10] considered the finite size of individual cells-"volume-filling'' and derived volume-filling models

    {ut=(Du(q(u)q(u)u)uq(u)uχ(v)v)+f(u,v),vt=DvΔv+g(u,v). (1.4)

    q(u) is the probability of the cell finding space at its neighbouring location. It is also called the squeezing probability, which reflects the elastic properties of cells. For the linear choice of q(u)=1u, global existence of solutions to model (1.4) in any space dimension are investigated in [9]. Wang and Thomas [11] established the global existence of classical solutions and given necessary and sufficient conditions for spatial pattern formation to a generalized volume-filling chemotaxis model. For a chemotaxis system with generalized volume-filling effect and logistic source, the global boundedness and finite time blow-up of solutions are obtained in [12]. Furthermore, the pattern formation of the volume-filling chemotaxis systems with logistic source and both linear diffusion and nonlinear diffusion are shown in [13,14,15] by the weakly nonlinear analysis. For parabolic-elliptic Keller-Segel volume-filling chemotaxis model with linear squeezing probability, asymptotic behavior of solutions is studied both in the whole space Rn [16] and on bounded domains [17]. Moreover, the boundedness and singularity formation in parabolic-elliptic Keller-Segel volume-filling chemotaxis model with nonlinear squeezing probability are discussed in [18,19].

    Very recently, we [20] investigated the uniform boundedness and global asymptotic stability for the following chemotaxis model of multiple sclerosis

    {ut=Δu(χ(u)v)+u(1u),χ(u)=χu1+u,xΩ,t>0,τvt=Δvβv+αw+γu,xΩ,t>0,wt=δu1+uu(1w),xΩ,t>0,

    subject to the homogeneous Neumann boundary conditions.

    In this paper, we are first devoted to studying the local existence and uniform boundedness of the unique classical solution to system (1.3) by using Neumann heat semigroup arguments, Banach fixed point theorem, parabolic Schauder estimate and elliptic regularity theory. Then we discuss that exponential asymptotic stability of the positive equilibrium point to system (1.3) by constructing Lyapunov function.

    Although, in the pathological mechanism of BCS, the initial data in model (1.3) satisfy 0<u0(x)1,w0(x)=0, we mathematically assume that

    {u0(x)C0(ˉΩ)with0,u0(x)1inΩ,w0(x)C2+ν(ˉΩ)with0<ν<1and0w0(x)1inΩ. (1.5)

    It is because the condition (1.5) implies u(x,t0)>0 for any t0>0 by the strong maximum principle.

    The following theorems give the main results of this paper.

    Theorem 1.1. Assume that the initial data (u0(x),w0(x)) satisfy the condition (1.5). Then model (1.3) possesses a unique global solution (u(x,t),v(x,t),w(x,t)) satisfying

    u(x,t)C0(ˉΩ×[0,))C2,1(ˉΩ×(0,)),v(x,t)C0((0,),C2(ˉΩ)),w(x,t)C2,1(ˉΩ×[0,)), (1.6)

    and

    0<u(x,t)1,0v(x,t)β+γα,w0(x)w(x,t)1,inˉΩ×(0,).

    Moreover, there exist a ν(0,1) and M>0 such that

    (1.7)

    Theorem 1.2. Assume that \beta \geq 0, \; \gamma\geq 0, \; \beta+\gamma > 0 and

    \begin{equation} \chi \lt \left\{\begin{array}{ll} \min \left\{\frac{2\sqrt{2\alpha}}{\beta}, \frac{2\sqrt{2\alpha}}{\gamma}\right\}, \, &\beta \gt 0, \gamma \gt 0, \\ \frac{2\sqrt{2\alpha}}{\beta}, \, &\beta \gt 0, \gamma = 0, \\ \frac{2\sqrt{2\alpha}}{\gamma}, \, &\beta = 0, \gamma \gt 0.\end{array} \right. \end{equation} (1.8)

    Let (u, v, w) be a positive classical solution of the problem (1.3), (1.5). Then

    \begin{equation} \|u(\cdot, t)-u^{\ast}\|_{L^{\infty}(\Omega)}+\|v(\cdot, t)-v^{\ast}\|_{L^{\infty}(\Omega)} +\|w(\cdot, t)-w^{\ast}\|_{L^{\infty}(\Omega)}\rightarrow 0, \; \; \mathit{\text{as}}\, t\rightarrow \infty. \end{equation} (1.9)

    Furthermore, there exist positive constants \lambda = \lambda(\chi, \alpha, \gamma, \delta, n) and C = C(|\Omega|, \chi, \alpha, \beta, \gamma, \delta) such that

    \begin{equation} \|u-u^{\ast}\|_{L^{\infty}(\Omega)}\leq C e^{-\lambda t}, \, \|v-v^{\ast}\|_{L^{\infty}(\Omega)}\leq C e^{-\lambda t}, \, \|w-w^{\ast}\|_{L^{\infty}(\Omega)} \leq C e^{-\lambda t}, \; \; t \gt 0, \end{equation} (1.10)

    where (u^{\ast}, v^{\ast}, w^{\ast}) = (1, \frac{\beta+\gamma}{\alpha}, 1) is the unique positive equilibrium point of the model (1.3).

    The paper is organized as follows. In section 2, we prove the local existence, the boundedness and global existence of a unique classical solution. In section 3, we firstly establish the uniform convergence of the positive global classical solution, then discuss the exponential asymptotic stability of positive equilibrium point in the case of weak chemotactic sensitivity. The paper ends with a brief concluding remarks.

    The aim of this section is to develop the existence and boundedness of a global classical solution by employing Neumann heat semigroup arguments, Banach fixed point theorem, parabolic Schauder estimate and elliptic regularity theory.

    Proof of Theorem 1.1 (ⅰ) Existence. For p\in (1, \infty) , let A denote the sectorial operator defined by

    Au: = -\Delta u \; \mathrm{for}\; u\in D(A): = \Big\{\varphi\in W^{2, p}(\Omega)\Big|\frac{\partial}{\partial \eta}\varphi\Big|_{\partial\Omega} = 0\Big\}.

    \lambda_{1} > 0 denote the first nonzero eigenvalue of -\Delta in \Omega with zero-flux boundary condition. Let A_{1} = -\Delta+\alpha and X^{l} be the domains of fractional powers operator A^{l}, \; l\geq 0 . From the Theorem 1.6.1 in [21], we know that for any p > n and l\in(\frac{n}{2p}, \frac{1}{2}) ,

    \begin{equation} \|z\|_{L^{\infty}(\Omega)}\leq C\|A_{1}^{l}z\|_{ L^{p}(\Omega)}\, \, \mathrm{for\; all}\, \, z\in X^{l}. \end{equation} (2.1)

    We introduce the closed subset

    S: = \left\{u\in X\big| \|u\|_{L^{\infty}((0, T);L^{\infty}(\Omega))}\leq R+1\right\}

    in the space X: = C^{0}([0, T];C^{0}(\bar{\Omega})) , where R is a any positive number satisfying

    \|u_{0}(x)\|_{L^{\infty}(\Omega)}\leq R

    and T > 0 will be specified later. Note F(u) = \frac{u}{1+u} , we consider an auxiliary problem with F(u) replaced by its extension \tilde{F}(u) defined by

    \tilde{F}(u) = \begin{cases} F(u)u\; \; &\text{if}\; \; u\geq 0, \\ -F(-u)(-u)\; \; &\text{if}\; \; u \lt 0. \end{cases}

    Notice that \tilde{F}(u) is a smooth globally Lipshitz function. Given \hat{u}\in S , we define \Psi\hat{u} = u by first writing

    \begin{equation} w(x, t) = (w_{0}(x)-1)e^{-\delta\int_{0}^{t}\tilde{F} (\hat{u})\hat{u}ds}+1, \; \; x\in\Omega, \; t \gt 0, \end{equation} (2.2)

    and

    w_{0}\leq w(x, t)\leq 1, \; \; x\in\Omega, \; t \gt 0,

    then letting v solve

    \begin{equation} \left\{\begin{array}{ll} -\Delta v+\alpha v = \beta w+\gamma \hat{u}, &x\in\Omega, \; t\in(0, T), \\ \partial_{\eta}v = 0, &x\in\partial\Omega, \; t\in(0, T), \\ \end{array} \right. \end{equation} (2.3)

    and finally taking u to be the solution of the linear parabolic problem

    \left\{\begin{array}{ll} u_{t} = \Delta u-\chi \nabla\cdot(\hat{u}(1-\hat{u})\nabla v)+\hat{u}(1-\hat{u}), &x\in\Omega, \; t\in(0, T), \\ \partial_{\eta}u = 0, &x\in\partial\Omega, \; t\in(0, T), \\ u(x, 0) = u_{0}(x), &x\in\Omega.\\ \end{array} \right.

    Applying Agmon-Douglas-Nirenberg Theorem [22,23] for the problem (2.3), there exists a constant C such that

    \begin{equation} \begin{aligned} \|v\|_{W_{p}^{2}(\Omega)}&\leq C(\beta\|w\|_{L^{p}(\Omega)}+\gamma\|\hat{u}\|_{L^{p}(\Omega)})\\ &\leq C(\beta|\Omega|^{\frac{1}{p}}+\gamma (R+1)) \end{aligned} \end{equation} (2.4)

    for all t\in(0, T) . From a variation-of-constants formula, we define

    \Psi(\hat{u}) = e^{t\Delta}u_{0}-\chi\int^{t}_{0}e^{(t-s)\Delta}\nabla\cdot\left(\hat{u}(1-\hat{u})\nabla v(s)\right)ds+\int^{t}_{0}e^{(t-s)\Delta}\hat{u}(s)(1-\hat{u}(s))ds.

    First we shall show that for T small enough

    \|\Psi(\hat{u})\|_{L^{\infty}((0, T);L^{\infty}(\Omega))}\leq R+1

    for any \hat{u}\in S . From the maximum principle, we can give

    \begin{equation} \|e^{t\Delta}u_{0}\|_{L^{\infty}(\Omega)}\leq \|u_{0}\|_{L^{\infty}(\Omega)}, \end{equation} (2.5)

    and

    \begin{equation} \begin{aligned} \int^{t}_{0}\|e^{t\Delta}\hat{u}(s)(1-\hat{u}(s))\|_{L^{\infty} (\Omega)}ds \leq& \int^{t}_{0}\|\hat{u}(s)(1-\hat{u}(s))\|_{L^{\infty} (\Omega)}ds\\ \leq&(R+1)(R+2)T \end{aligned} \end{equation} (2.6)

    for all t\in(0, T) . We use inequalities (2.1) and (2.4) to estimate

    \begin{equation} \begin{aligned} &\chi\int^{t}_{0}\|e^{(t-s)\Delta}\nabla\cdot(\hat{u}(1-\hat{u})\nabla v(s))\|_{L^{\infty}(\Omega)}ds\\ \leq& C\int^{t}_{0}(t-s)^{-l} \|e^{\frac{t-s}{2}\Delta}\nabla\cdot(\hat{u}(1-\hat{u})\nabla v(s))\|_{L^{p}(\Omega)}ds \\ \leq& C\int^{t}_{0}(t-s)^{-l-\frac{1}{2}} \|(\hat{u}(1-\hat{u})\nabla v(s)\|_{L^{p}(\Omega)}ds \\ \leq& C T^{\frac{1}{2}-l}(R+1)(R+2)(\beta|\Omega|^{\frac{1}{p}}+\gamma (R+1)) \end{aligned} \end{equation} (2.7)

    for all t\in(0, T) . This estimate is attributed to T < 1 and the inequality in [24], Lemma 1.3 iv]

    \| e^{t\Delta}\nabla z\|_{L^{p}(\Omega)}\leq C_{1}(1+t^{-\frac{1}{2}})e^{-\lambda_{1}t}\| z\|_{L^{p}(\Omega)}\; \mathrm{for\; all}\; \; z\in C^{\infty}_{c}(\Omega).

    From inequalities (2.5), (2.6) and (2.7) we can deduce that \Psi maps S into itself for T small enough.

    Next we prove that the map \Psi is a contractive on S . For \hat{u}_{1}, \hat{u}_{2}\in S , we estimate

    \begin{align*} \; \; \; \; \; \; \; \; &\|\Psi(\hat{u}_{1})-\Psi(\hat{u}_{2})\|_{L^{\infty}(\Omega)} \\ \leq & \chi \int^{t}_{0}(t-s)^{-l-\frac{1}{2}}\|\left[\hat{u}_{2}(s)(1-\hat{u}_{2}(s))-\hat{u}_{1}(s)(1-\hat{u}_{1}(s))\right] \nabla v_{2}(s)\|_{L^{p}(\Omega)}ds\\ &+\chi \int^{t}_{0}\|\hat{u}_{1}(s)(1-\hat{u}_{1}(s))(\nabla v_{1}(s)- \nabla v_{2}(s))\|_{L^{p}(\Omega)}ds \\ &+\int^{t}_{0}\|e^{(t-s)\Delta}[\hat{u}_{1}(s)(1-\hat{u}_{1}(s))-\hat{u}_{2}(s)(1-\hat{u}_{2}(s))]\|_{L^{\infty}(\Omega)}ds \\ \leq & \chi \int^{t}_{0}(t-s)^{-l-\frac{1}{2}}(2R+1)\|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{X}\|\nabla v_{2}(s)\|_{L^{p}(\Omega)}ds\\ &+\chi \int^{t}_{0}(R+1)(R+2)\left(\beta \|w_{1}(s)-w_{2}(s)\|_{L^{p}(\Omega)}+\gamma \|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{L^{p}(\Omega)}\right)ds \\ & +\int^{t}_{0}(2R+1)\|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{X}ds \\ \leq & \chi \int^{t}_{0}(t-s)^{-l-\frac{1}{2}}(2R+1)\|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{X}\|\nabla v_{2}(s)\|_{L^{p}(\Omega)}ds\\ &+2\beta\delta \chi \int^{t}_{0}(R+1)(R+2)t\|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{L^{p}(\Omega)}+\gamma \|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{L^{p}(\Omega)}ds\\ & +\int^{t}_{0}(2R+1)\|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{X}ds \\ \leq & \left(C\chi T^{\frac{1}{2}-l}(2R+1)(\beta|\Omega|^{\frac{1}{p}}+\gamma (R+1))+2\beta\delta \chi T(R^{2}+3R+\gamma+2)+T(2R+1)\right)\|\hat{u}_{1}(s)-\hat{u}_{2}(s)\|_{X}. \end{align*}

    Fixing T\in(0, 1) small enough such that

    \left(C\chi T^{\frac{1}{2}-l}(2R+1)(\beta|\Omega|^{\frac{1}{p}}+\gamma (R+1))+2\beta\delta \chi T(R^{2}+3R+\gamma+2)+T(2R+1)\right)\leq \frac{1}{2}.

    It follows from the Banach fixed point theorem that there exists a unique fixed point of \Psi .

    (ⅱ) Regularity. Since the above of T depends on \|u_{0}\|_{L^{\infty}(\Omega)} and \|w_{0}\|_{L^{\infty}(\Omega)} only, it is clear that (u, v, w) can be extended up to some maximal T_{\max}\in(0, \infty] . Let Q_{T} = \Omega \times (0, T] for all T\in (0, T_{\max}) . From u\in C^{0}(\bar{Q}_{T}) , we know that w\in C^{0, 1}(\bar{Q}_{T}) by the expression (2.2) and v\in C^{0}([0, T], W_{p}^{2}(\Omega)) by Agmon-Douglas-Nirenberg Theorem [22,23]. From parabolic L^{p} -estimate and the embedding relation W_{p}^{1}(\Omega)\hookrightarrow C^{\nu}(\bar{\Omega}), \; p > n , we can get u\in W^{2, 1}_{p}(Q_{T}) . By applying the following embedding relation

    \begin{equation} W^{2, 1}_{p}(Q_{T})\hookrightarrow C^{\nu, \nu/2}(\bar{Q}_{T}), \; p \gt \frac{n+2}{2}, \end{equation} (2.8)

    we can derive u(x, t)\in C^{\nu, \nu/2}(\bar{Q}_{T}) with 0 < \nu\leq 2-\frac{n+2}{p} . The conclusion w\in C^{\nu, 1+\nu/2}(\bar{Q}_{T}) can be obtained by substituting u\in C^{\nu, \nu/2}(\bar{Q}_{T}) into the formulation (2.2). The regularity u\in C^{2+\nu, 1+\nu/2}(\bar{Q}_{T}) can be deduced by using further bootstrap argument and the parabolic Schauder estimate. Similarly, we can get v\in C^{0}((0, T), C^{2+\nu}(\bar{\Omega})) by using Agmon-Douglas-Nirenberg Theorem [22,23]. From the regularity of u we have w\in C^{2+\nu, 1+\nu/2}(\bar{Q}_{T}) .

    Moreover, the maximal principle entails that 0 < u(x, t)\leq 1 , 0\leq v(x, t)\leq\frac{\beta+\gamma}{\alpha} . It follows from the positivity of u that \tilde{F}(u) = F(u) and because of the uniqueness of solution we infer the existence of the solution to the original problem.

    (ⅲ) Uniqueness. Suppose (u_{1}, v_{1}, w_{1}) and (u_{2}, v_{2}, w_{2}) are two deferent solutions of model (1.3) in \Omega\times [0, T] . Let U = u_{1}-u_{2} , V = v_{1}-v_{2} , W = w_{1}-w_{2} for t\in (0, T) . Then

    \begin{equation} \begin{aligned} &\frac{1}{2}\frac{d}{dt}\int_{\Omega}U^{2}dx+\int_{\Omega}|\nabla U|^{2}dx\\ \leq& \chi\int_{\Omega}|u_{1}(1-u_{1})-u_{2}(1-u_{2})|\nabla v_{1}||\nabla U|+u_{2}(1-u_{2})|\nabla V||\nabla U| dx\\ &+\int_{\Omega}|u_{1}(1-u_{1})-u_{2}(1-u_{2})||U|dx \\ \leq & \chi\int_{\Omega}|U||\nabla v_{1}||\nabla U|+\frac{1}{4}|\nabla V||\nabla U| dx +\int_{\Omega}|U|^{2}dx \\ \leq &\int_{\Omega}|\nabla U|^{2}dx+\frac{\chi^{2}}{32}\int_{\Omega}|\nabla V|^{2}dx+ \frac{\chi^{2} K^{2}+2}{2}\int_{\Omega}|U|^{2}dx, \end{aligned} \end{equation} (2.9)

    where we have used that |\nabla v_{1}|\leq K results from \nabla v_{1}\in C^{0}([0, T], C^{0}(\bar{\Omega})).

    Similarly, by Young inequality and w_{0}\leq w_{1}\leq 1 , we can estimate

    \begin{equation} \int_{\Omega}|\nabla V|^{2}dx+\frac{\alpha}{2}\int_{\Omega}| V|^{2}dx\leq\frac{\beta^{2}}{\alpha} \int_{\Omega}|W|^{2}dx+\frac{\gamma^{2}}{\alpha} \int_{\Omega}|U|^{2}dx, \end{equation} (2.10)

    and

    \begin{equation} \frac{d}{dt}\int_{\Omega}W^{2}dx\leq \delta\int_{\Omega}|U|^{2}+|W|^{2}dx. \end{equation} (2.11)

    Finally, adding to the inequalities (2.9)–(2.11) yields

    \frac{d}{dt}\left(\int_{\Omega}U^{2}dx+\int_{\Omega}W^{2}dx\right)\leq C\left(\int_{\Omega}U^{2}dx+\int_{\Omega}W^{2}dx\right)\; \mathrm{for\; all}\; t \in (0, T).

    The results U\equiv 0 , W\equiv0 in \Omega\times(0, T) are obtained by Gronwall's lemma. From the inequality (2.10), we have V\equiv 0 . Hence (u_{1}, v_{1}, w_{1}) = (u_{2}, v_{2}, w_{2}) in \Omega\times(0, T) .

    (ⅳ) Uniform estimates. We use the Agmon-Douglas-Nirenberg Theorem [22,23] for the second equation of the model (1.3) to get

    \begin{equation} \|v\|_{C^{0}([t, t+1], W_{p}^{2}(\Omega))}\leq C\left(\|u\|_{L^{p}(\Omega \times [t, t+1])}+\|w\|_{L^{p}(\Omega \times [t, t+1])}\right) \leq C_{2} \end{equation} (2.12)

    for all t\geq 1 and C_{2} is independent of t . From the embedded relationship W_{p}^{1}(\Omega)\hookrightarrow C^{0}({\bar{\Omega}}), \; p > n , the parabolic L^{p} -estimate and the estimation (2.12), we have

    \|u\|_{W_{p}^{2, 1}(\Omega\times[t, t+1])}\leq C_{3}

    for all t\geq 1 . The estimate \|u\|_{C^{\nu, \frac{\nu}{2}}(\bar{\Omega}\times [t, t+1])}\leq C_{4} for all t\geq 1 obtained by the embedded relationship (2.8). We can immediately compute \|w\|_{C^{\nu, 1+\frac{\nu}{2}}(\bar{\Omega}\times [t, t+1])}\leq C_{5} for all t\geq 1 according to the regularity of u and the specific expression of w . Further, bootstrapping argument leads to \|v\|_{C^{0}([t, t+1], C^{2+\nu}(\bar{\Omega}))}\leq C_{6} and \|u\|_{C^{2+\nu, 1+\frac{\nu}{2}}(\bar{\Omega}\times [t, t+1])}\leq C_{7} for all t\geq 1 . Thus the uniform estimation (1.7) is proved.

    Remark 2.1. Assume the initial data 0 < u_{0}(x)\leq 1 and w_{0}(x) = 0 . Then the BCS model (1.3) has a unique classical solution.

    In this section we investigate the global asymptotic stability of the unique positive equilibrium point (1, \frac{\beta+\gamma}{\alpha}, 1) to model (1.3). To this end, we first introduce following auxiliary problem

    \begin{equation} \left\{\begin{array}{ll} u_{\epsilon t} = \Delta u_{\epsilon}-\nabla\cdot(u_{\epsilon}(1-u_{\epsilon})\nabla v_{\epsilon})+u_{\epsilon}(1-u_{\epsilon}), & x\in\Omega, \; t \gt 0, \\ -\Delta v_{\epsilon}+\alpha v_{\epsilon} = \beta w_{\epsilon}+\gamma u_{\epsilon}, &x\in\Omega, \; t \gt 0, \\ w_{\epsilon t} = \delta\frac{u_{\epsilon}^{2}+\epsilon}{1+u_{\epsilon}} (1-w_{\epsilon}), &x\in\Omega, \; t \gt 0, \\ \partial_{\eta}u_{\epsilon} = \partial_{\eta}v_{\epsilon} = 0, &x\in\partial\Omega, \; t \gt 0, \\ u_{\epsilon}(x, 0) = u_{0}(x), \; w_{\epsilon}(x, 0) = w_{0}(x), &x\in\Omega. \end{array} \right. \end{equation} (3.1)

    By a similar proof of Theorem 1.1, we get that the problem (3.1) has a unique global classical solution (u_{\epsilon}, v_{\epsilon}, w_{\epsilon}) , and there exist a \nu\in(0, 1) and M_{1} > 0 which is independent of \epsilon such that

    \begin{equation} \|u_{\epsilon}\|_{C^{2+\nu, 1+\nu/2}(\bar{\Omega}\times[1, \infty))}+\|v_{\epsilon}\|_{C^{2+\nu, 1+\nu/2}(\bar{\Omega}\times[1, \infty))} +\|w_{\epsilon}\|_{C^{\nu, 1+\nu/2}(\bar{\Omega}\times[1, \infty))}\leq M_{1}. \end{equation} (3.2)

    Then, motivated by some ideas from [25,26], we construct a Lyapunov function to study the uniform convergence of homogeneous steady state for the problem (3.1).

    Let us give following lemma which is used in the proof of Lemma 3.2.

    Lemma 3.1. Suppose that a nonnegative function f on (1, \infty) is uniformly continuous and \int_{1}^{\infty}f(t)dt < \; \infty . Then f(t)\rightarrow 0 as t\rightarrow \infty.

    Lemma 3.2. Assume that the condition (1.8) is satisfied. Then

    \begin{equation} \|u_{\epsilon}(\cdot, t)-1\|_{L^{2}(\Omega)}+ \|v_{\epsilon}(\cdot, t)-v^{\ast}\|_{L^{2}(\Omega)} +\|w_{\epsilon}(\cdot, t)-1\|_{L^{2}(\Omega)}\rightarrow 0, \; \; t\rightarrow \infty, \end{equation} (3.3)

    where v^{\ast} = \frac{\beta+\gamma}{\alpha} .

    Proof We construct a positive function

    E(t): = \int_{\Omega}(u_{\varepsilon}-1-\ln u_{\epsilon}) +\frac{1}{2\delta\epsilon}\int_{\Omega}(w_{\epsilon}-1)^{2}, \; \; t \gt 0.

    From the problem (3.1) and Young's inequality, we can compute

    \begin{equation} \frac{d}{dt}E(t)\leq {\frac{\chi^{2}}{4}}\int_{\Omega}|\nabla v_{\epsilon}|^{2}dx-\int_{\Omega}(u_{\epsilon}-1)^{2}dx-\int_{\Omega}(w_{\epsilon}-1)^{2}dx, \; \; t \gt 0. \end{equation} (3.4)

    We multiply the second equations in system (3.1) by v_{\epsilon}-v^{\ast} , integrate by parts over \Omega and use Young's inequality to obtain

    \begin{equation} \int_{\Omega}|\nabla v_{\epsilon}|^{2}dx\leq\frac{\gamma^{2}}{2\alpha}\int_{\Omega}(u_{\epsilon}-1)^{2}dx +\frac{\beta^{2}}{2\alpha}\int_{\Omega}(w_{\epsilon}-1)^{2}dx, \; \; t \gt 0, \end{equation} (3.5)

    and

    \begin{equation} \int_{\Omega}(v_{\epsilon}-v^{\ast})^{2}dx\leq\frac{2\gamma^{2}}{\alpha^{2}}\int_{\Omega}(u_{\epsilon}-1)^{2}dx+\frac{2 \beta^{2}}{\alpha^{2}}\int_{\Omega}(w_{\epsilon}-1)^{2}dx, \; \; t \gt 0. \end{equation} (3.6)

    Substituting inequality (3.5) into inequality (3.4) to get

    \begin{equation} \nonumber \frac{d}{dt}E(t)\leq -C_{8}\left(\int_{\Omega}(u_{\epsilon}-1)^{2}dx+\int_{\Omega}(w_{\epsilon}-1)^{2}dx\right), \; \; t \gt 0, \end{equation}

    where C_{8} = \min\left\{1-\frac{\chi^{2}\beta^{2}}{8\alpha}, 1-\frac{\chi^{2}\gamma^{2}}{8\alpha}\right\} > 0.

    Let f(t): = \int_{\Omega}(u_{\epsilon}-1)^{2}+(w_{\epsilon}-1)^{2}dx . Then

    \int_{1}^{\infty}f(t)dt\leq \frac{E(1)}{C_{8}} \lt \infty, \; \; t \gt 1.

    It follows from the uniform estimation (3.2) and the Arzela-Ascoli theorem that f(t) is uniformly continuous in (1, \infty) . Applying Lemma 3.1, we have

    \begin{equation} \int_{\Omega}(u_{\epsilon}(\cdot, t)-1)^{2}+ (w_{\epsilon}(\cdot, t)-1)^{2}dx\rightarrow 0, \; \; t\rightarrow \infty. \end{equation} (3.7)

    Combining inequality (3.6) and the limit (3.7) to obtain

    \int_{\Omega}(v_{\epsilon}(\cdot, t)-v^{\ast})^{2}dx \rightarrow 0, \; \; t\rightarrow \infty.

    Proof of Theorem 1.2 As we all known, each bounded sequence in C^{2+\nu, 1+\frac{\nu}{2}}(\bar{\Omega}\times[1, \infty)) is precompact in C^{2, 1}(\bar{\Omega}\times[1, \infty)) . Hence there exists some subsequence \{u_{\epsilon_{n}}\}_{n = 1}^{\infty} satisfying \epsilon_{n}\rightarrow0 as n\rightarrow \infty such that

    \lim\limits_{n\rightarrow \infty}\|u_{\epsilon_{n}}-u_{\ast}\|_{C^{2, 1}(\bar{\Omega}\times[1, \infty))} = 0.

    Similarly, we can get

    \lim\limits_{n\rightarrow \infty}\|v_{\epsilon_{n}}-v_{\ast}\|_{C^{2}(\bar{\Omega})} = 0,

    and

    \lim\limits_{n\rightarrow \infty}\|w_{\epsilon_{n}}-w_{\ast}\|_{C^{0, 1}(\bar{\Omega}\times[1, \infty))} = 0.

    Combining above limiting relations yields that (u_{\ast}, v_{\ast}, w_{\ast}) satisfies model (1.3). The conclusion (u_{\ast}, v_{\ast}, w_{\ast}) = (u, v, w) is directly attributed to the uniqueness of the classical solution of the model (1.3). Furthermore, according to the conclusion, the strong convergence (3.3) and Diagonal line method, we can deduce

    \begin{equation} \|u(\cdot, t)-1\|_{L^{2}(\Omega)}+ \|v(\cdot, t)-v^{\ast}\|_{L^{2}(\Omega)} +\|w(\cdot, t)-1\|_{L^{2}(\Omega)}\rightarrow 0, \; \; t\rightarrow \infty. \end{equation} (3.8)

    By applying Gagliardo-Nirenberg inequality

    \begin{equation} \|z\|_{L^{\infty}}\leq C\|z\|_{L^{2}(\Omega)}^{2/(n+2)}\|z\|_{W^{1, \infty}(\Omega)}^{n/(n+2)}, \; \; z\in W^{1, \infty}(\Omega), \end{equation} (3.9)

    comparison principle of ODE and the convergence (3.8), the uniform convergence (1.9) is obtained immediately.

    Since \lim_{t\rightarrow \infty}\|u(\cdot, t)-1\|_{L^{\infty}(\Omega)} = 0 , so there exists a t_{1} > 0 such that

    \begin{equation} u(x, t)\geq \frac{1}{2}\; \; \mathrm{for\; all}\; \; x\in \Omega, \; \; t \gt t_{1}. \end{equation} (3.10)

    Using the explicit representation formula of w

    w(x, t) = (w_{0}(x)-1)e^{-\delta\int_{0}^{t}F(u)uds}+1, \; \; x\in\Omega, \; t \gt 0

    and the inequality (3.10), we have

    \begin{equation} \|w(\cdot, t)-1\|_{L^{\infty}(\Omega)}\leq e^{-\frac{\delta}{6}(t-t_{1})}, \; \; t \gt t_{1}. \end{equation} (3.11)

    Multiply the first two equations in model (1.3) by u-1 and v-v^{\ast} , respectively, integrate over \Omega and apply Cauchy's inequality, Young's inequality and the inequality (3.10), to find

    \begin{equation} \frac{d}{dt}\int_{\Omega}(u-1)^{2}dx\leq \frac{\chi^{2}}{32}\int_{\Omega}|\nabla v|^{2}dx-\int_{\Omega}(u-1)^{2}dx, \; \; t \gt t_{1}. \end{equation} (3.12)
    \begin{equation} \int_{\Omega}|\nabla v|^{2}dx+\frac{\alpha}{2}\int_{\Omega}(v-v^{\ast})^{2}dx\leq \frac{\beta^{2}}{\alpha}\int_{\Omega}(w-1)^{2}dx +\frac{\gamma^{2}}{\alpha}\int_{\Omega}(u-1)^{2}dx, \; \; t \gt 0. \end{equation} (3.13)

    Combining the estimations (3.11)–(3.13) leads us to the estimate

    \begin{equation} \nonumber \frac{d}{dt}\int_{\Omega}(u-1)^{2}dx\leq \left(\frac{\chi^{2}\gamma^{2}}{32\alpha}-1\right)\int_{\Omega}(u-1)^{2}dx +\frac{\chi^{2}\beta^{2}}{32\alpha}e^{-{\frac{\delta}{3}(t-t_{1})}}, \; \; t \gt t_{1}. \end{equation}

    Let y(t) = \int_{\Omega}(u-1)^{2}dx . Then

    y'(t)\leq \left(\frac{\chi^{2}\gamma^{2}}{32\alpha}-1\right)y(t) +\frac{\chi^{2}\beta^{2}}{32\alpha}e^{-{\frac{\delta}{3}(t-t_{1})}}, \; \; t \gt t_{1}.

    From comparison principle of ODE, we get

    y(t)\leq \left(y(t_{1})-\frac{3\chi^{2}\beta^{2}}{32\alpha(3-\delta)-\chi^{2}\gamma^{2}}\right) e^{-\left(1-\frac{\chi^{2}\gamma^{2}}{32\alpha}\right)(t-t_{1})} +\frac{3\chi^{2}\beta^{2}}{32\alpha(3-\delta)-\chi^{2}\gamma^{2}}e^{-\frac{\delta}{3}(t-t_{1})}, \; \; t \gt t_{1}.

    This yields

    \begin{equation} \int_{\Omega}(u-1)^{2}dx\leq C_{9} e^{-\lambda_{2} (t-t_{1})}, \; \; t \gt t_{1}, \end{equation} (3.14)

    where \lambda_{2} = \min\{1-\frac{\chi^{2}\gamma^{2}}{32\alpha}, \frac{\delta}{3}\} and C_{9} = \max\left\{|\Omega|-\frac{3\chi^{2}\beta^{2}}{32\alpha(3-\delta)-\chi^{2}\gamma^{2}}, \frac{3\chi^{2}\beta^{2}}{32\alpha(3-\delta)-\chi^{2}\gamma^{2}}\right\} .

    From the inequalities (3.11), (3.13) and (3.14), we derive

    \begin{equation} \int_{\Omega}\left(v-\frac{\beta+\gamma}{\alpha}\right)^{2}dx\leq C_{10}e^{-\lambda_{2}(t-t_{1})}, \; \; t \gt t_{1}, \end{equation} (3.15)

    where C_{10} = \max\left\{\frac{2\gamma^{2}}{\alpha^{2}}C_{9}, \frac{2\beta^{2}}{\alpha^{2}}\right\} . By employing the uniform estimation (1.7), the inequalities (3.9), (3.14) and (3.15), the exponential decay estimation (1.10) can be obtained.

    The proof is complete.

    In this paper, we mainly study the uniform boundedness of classical solutions and exponential asymptotic stability of the unique positive equilibrium point to the chemotactic cellular model (1.3) for Baló's concentric sclerosis (BCS). For model (1.1), by numerical simulation, Calveza and Khonsarib in [7] shown that demyelination patterns of concentric rings will occur with increasing of chemotactic sensitivity. By the Theorem 1.1 we know that systems (1.1) and (1.2) are {uniformly} bounded and dissipative. By the Theorem 1.2 we also find that the constant equilibrium point of model (1.1) is exponentially asymptotically stable if

    \tilde{\chi} \lt \frac{2}{\bar{w}\tilde{\beta}} \sqrt{\frac{2D\mu\tilde{\alpha}\tilde{\epsilon}}{\bar{u}}},

    and the constant equilibrium point of the model (1.2) is exponentially asymptotically stable if

    \tilde{\chi} \lt 2\sqrt{\frac{2D\mu\tilde{\alpha}\tilde{\epsilon}}{\bar{u}}}\min \left\{\frac{1}{\bar{w}\tilde{\beta}}, \frac{1}{\bar{u}\tilde{\gamma}}\right\}.

    According to a pathological viewpoint of BCS, the above stability results mean that if chemoattractive effect is weak, then the destroyed oligodendrocytes form a homogeneous plaque.

    The authors would like to thank the editors and the anonymous referees for their constructive comments. This research was supported by the National Natural Science Foundation of China (Nos. 11761063, 11661051).

    We have no conflict of interest in this paper.



    [1] P. Voigt, A. Von dem Bussche, The eu general data protection regulation (gdpr), A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10 (2017), 5510–5555. https://doi.org/10.1007/978-3-319-57959-7 doi: 10.1007/978-3-319-57959-7
    [2] G. J. Annas, Hipaa regulations: a new era of medical-record privacy?, New England J. Med., 348 (2003), 1486. https://doi.org/10.1056/NEJMlim035027 doi: 10.1056/NEJMlim035027
    [3] A. Hard, K. Rao, R. Mathews, S. Ramaswamy, F. Beaufays, S. Augenstein, Het al., Federated learning for mobile keyboard prediction, preprint, arXiv: 1811.03604.
    [4] N. Rieke, J. Hancox, W. Li, F. Milletari, H. R. Roth, S. Albarqouni, et al., The future of digital health with federated learning, NPJ Dig. Med., 3 (2020), 119. https://doi.org/10.1038/s41746-020-00323-1 doi: 10.1038/s41746-020-00323-1
    [5] X. Wang, S. Garg, H. Lin, G. Kaddoum, J. Hu, M. S. Hossain, A secure data aggregation strategy in edge computing and blockchain-empowered internet of things, IEEE Int. Things J., 9 (2020), 14237–14246. https://doi.org/10.1109/JIOT.2020.3023588 doi: 10.1109/JIOT.2020.3023588
    [6] X. Wang, S. Garg, H. Lin, G. Kaddoum, J. Hu, M. M. Hassan, Heterogeneous blockchain and ai-driven hierarchical trust evaluation for 5g-enabled intelligent transportation systems, IEEE Trans. Intell. Transp. Syst., 24 (2021), 2074–2083. https://doi.org/10.1109/TITS.2021.3129417 doi: 10.1109/TITS.2021.3129417
    [7] G. Long, Y. Tan, J. Jiang, C. Zhang, Federated learning for open banking, in Federated Learning: Privacy and Incentive, (2020), 240–254. https://doi.org/10.1007/978-3-030-63076-8_17
    [8] B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, Communication-efficient learning of deep networks from decentralized data, in Artificial intelligence and statistics, (2017), 1273–1282.
    [9] V. Smith, C. K. Chiang, M. Sanjabi, A. S. Talwalkar, Federated multi-task learning, in Advances in neural information processing systems, (2017).
    [10] F. Sattler, K. R. Müller, W. Samek, Clustered federated learning: Model-agnostic distributed multitask optimization under privacy constraints, IEEE Trans. Neural Networks Learning Syst., 32 (2020), 3710–3722. https://doi.org/10.1109/TNNLS.2020.3015958 doi: 10.1109/TNNLS.2020.3015958
    [11] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, V. Young, Mobile edge computing—a key technology towards 5g, ETSI White Paper, 11 (2015), 1–16.
    [12] E. Fazeldehkordi, T. M. Grønli, A survey of security architectures for edge computing-based iot, IoT, 3 (2022), 332–365. https://doi.org/10.3390/iot3030019 doi: 10.3390/iot3030019
    [13] K. Cao, Y. Liu, G. Meng, Q. Sun, An overview on edge computing research, IEEE Access, 8 (2020), 85714–85728. https://doi.org/10.1109/ACCESS.2020.2991734 doi: 10.1109/ACCESS.2020.2991734
    [14] K. B. Letaief, Y. Shi, J. Lu, J. Lu, Edge artificial intelligence for 6g: Vision, enabling technologies, and applications, IEEE J. Selected Areas Commun., 40 (2021), 5–36. https://doi.org/10.1109/JSAC.2021.3126076 doi: 10.1109/JSAC.2021.3126076
    [15] E. Kristiani, Y. T. Tsan, P. Y. Liu, N. Y. Yen, C. T. Yang, Binary and multi-class assessment of face mask classification on edge ai using cnn and transfer learning, Human Centric Comput. Inf. Sci., 12 (2022). https://doi.org/10.22967/HCIS.2022.12.053 doi: 10.22967/HCIS.2022.12.053
    [16] M. Babar, M. S. Khan, U. Habib, B. Shah, F. Ali, D. Song, Scalable edge computing for iot and multimedia applications using machine learning, Human centric Comput. Inf. Sci., 11 (2021). https://doi.org/10.22967/hcis.2021.11.041 doi: 10.22967/hcis.2021.11.041
    [17] B. He, T. Li, An offloading scheduling strategy with minimized power overhead for internet of vehicles based on mobile edge computing, J. Inf. Process. Syst., 17 (2021), 489–504. https://doi.org/10.3745/JIPS.01.0077 doi: 10.3745/JIPS.01.0077
    [18] Y. He, Z. Tang, Strategy for task offloading of multi-user and multi-server based on cost optimization in mobile edge computing environment, J. Inf. Process. Syst., 17 (2021), 615–629. https://doi.org/10.3745/JIPS.01.0078 doi: 10.3745/JIPS.01.0078
    [19] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, J. Zhang, Edge intelligence: Paving the last mile of artificial intelligence with edge computing, Proc. IEEE, 107 (2019), 1738–1762. https://doi.org/10.1109/JPROC.2019.2918951 doi: 10.1109/JPROC.2019.2918951
    [20] X. Zhu, H. Li, Y. Yu, Blockchain-based privacy preserving deep learning, in International Conference on Information Security and Cryptology, (2018), 370–383. https://doi.org/10.1007/978-3-030-14234-6_20
    [21] T. S. Brisimi, R. Chen, T. Mela, A. Olshevsky, I. C. Paschalidis, W. Shi, Federated learning of predictive models from federated electronic health records, Int. J. Med. Inf., 112 (2018), 59–67. https://doi.org/10.1016/j.ijmedinf.2018.01.007 doi: 10.1016/j.ijmedinf.2018.01.007
    [22] S. Samarakoon, M. Bennis, W. Saad, M. Debbah, Federated learning for ultra-reliable low-latency v2v communications, in 2018 IEEE Global Communications Conference (GLOBECOM), (2018), 1–7. https://doi.org/10.1109/GLOCOM.2018.8647927
    [23] M. Chen, R. Mathews, T. Ouyang, F. Beaufays, Federated learning of out-of-vocabulary words, preprint, arXiv: 1903.10635.
    [24] S. P. Karimireddy, S. Kale, M. Mohri, S. J. Reddi, S. U. Stich, A. T. Suresh, Scaffold: Stochastic controlled averaging for federated learning, in International Conference on Machine Learning, (2020), 5132–5143.
    [25] T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith, Federated optimization in heterogeneous networks, in Proceedings of Machine learning and systems, (2020), 429–450.
    [26] J. Wang, Q. Liu, H. Liang, G. Joshi, H. V. Poor, Tackling the objective inconsistency problem in heterogeneous federated optimization, in Advances in neural information processing systems, (2020), 7611–7623.
    [27] M. Firdaus, K. H Rhee, A joint framework to privacy-preserving edge intelligence in vehicular networks, in Information Security Applications: 23rd International Conference, WISA 2022, Cham: Springer Nature Switzerland, (2023), 156–167. https://doi.org/10.1007/978-3-031-25659-2_12
    [28] J. S. P. G. M. Nam, J. G. Shon, A blockchain-based cheating detection system for online examination, KIPS Trans. Software Data Eng., 11 (2022), 267–272. https://doi.org/10.3745/KTSDE.2022.11.6.267 doi: 10.3745/KTSDE.2022.11.6.267
    [29] A. Z. Tan, H. Yu, L. Cui, Q. Yang, Towards personalized federated learning, IEEE Trans. Neural Networks Learn. Syst., (2022), 1–17. https://doi.org/10.1109/TNNLS.2022.3160699 doi: 10.1109/TNNLS.2022.3160699
    [30] K. C. Sim, P. Zadrazil, F. Beaufays, An investigation into on-device personalization of end-to-end automatic speech recognition models, preprint, arXiv: 1909.06678.
    [31] V. Kulkarni, M. Kulkarni, A. Pant, Survey of personalization techniques for federated learning, in 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), (2020), 794–797. https://doi.org/10.1109/WorldS450073.2020.9210355
    [32] Y. Jiang, J. Konečnỳ, K. Rush, S. Kannan, Improving federated learning personalization via model agnostic meta learning, preprint, arXiv: 1909.12488.
    [33] Y. Zhao, M. Li, L. Lai, N. Suda, D. Civin, V. Chandra, Federated learning with non-iid data, preprint, arXiv: 1806.00582.
    [34] Y. Chen, X. Qin, J. Wang, C. Yu, W. Gao, Fedhealth: A federated transfer learning framework for wearable healthcare, IEEE Intell. Syst., 35 (2020), 83–93. https://doi.org/10.1109/MIS.2020.2988604 doi: 10.1109/MIS.2020.2988604
    [35] Y. Tian, T. Li, J. Xiong, M. Z. A. Bhuiyan, J. Ma, C. Peng, A blockchain-based machine learning framework for edge services in iiot, IEEE Trans. Ind. Inf., 18 (2021), 1918–1929. https://doi.org/10.1109/TII.2021.3097131 doi: 10.1109/TII.2021.3097131
    [36] J. Weng, J. Weng, J. Zhang, M. Li, Y. Zhang, W. Luo, Deepchain: Auditable and privacy-preserving deep learning with blockchain-based incentive, IEEE Trans. Dependable Secure Comput., 18 (2019), 2438–2455. https://doi.org/10.1109/TDSC.2019.2952332 doi: 10.1109/TDSC.2019.2952332
    [37] Y. Huang, L. Chu, Z. Zhou, L. Wang, J. Liu, J. Pei, et al., Personalized cross-silo federated learning on non-iid data, in Proceedings of the AAAI Conference on Artificial Intelligence, (2021), 7865–7873. https://doi.org/10.1609/aaai.v35i9.16960
    [38] X. Wang, S. Garg, H. Lin, M. J. Piran, J. Hu, M. S. Hossain, Enabling secure authentication in industrial iot with transfer learning empowered blockchain, IEEE Trans. Ind. Inf., 17 (2021), 7725–7733. https://doi.org/10.1109/TII.2021.3049405 doi: 10.1109/TII.2021.3049405
    [39] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proc. IEEE, 86 (1998), 2278–2324. https://doi.org/10.1109/5.726791 doi: 10.1109/5.726791
    [40] M. Firdaus, H. T. Larasati, K. H. Rhee, A secure federated learning framework using blockchain and differential privacy, in 2022 IEEE 9th International Conference on Cyber Security and Cloud Computing (CSCloud)/2022 IEEE 8th International Conference on Edge Computing and Scalable Cloud (EdgeCom), (2022), 18–23.
  • This article has been cited by:

    1. Lu Xu, Chunlai Mu, Qiao Xin, Global boundedness and asymptotic behavior of solutions for a quasilinear chemotaxis model of multiple sclerosis with nonlinear signal secretion, 2023, 28, 1531-3492, 1215, 10.3934/dcdsb.2022118
  • Reader Comments
  • © 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2773) PDF downloads(136) Cited by(8)

Figures and Tables

Figures(5)  /  Tables(1)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog