Federated Learning (FL) is a promising solution for distributed cooperative training in various scenarios. FL empowers data-rich devices to contribute to a comprehensive model, but it suffers from the data fairness problem, where local datasets are not fairly accounted for in the global model. This problem is especially devastating in the scenario of heterogeneous devices with non-lID data distributions. To resolve this issue, we propose Federated Learning with Data Fairness (DFFL), which incorporates model factorization to adapt local models of different sizes to devices of various computation capabilities, and a novel DNM metric for client selection to boost its performance on devices with minor data distributions. Experiments show that, the proposed DFFL outperforms existing FL schemes, achieving higher accuracy with much lower (82 % less local parameters on average) computation load, while providing a more robust assurance of data fairness among devices.

Federated Learning (FL) enables insurance companies to collaboratively train fraud detection models while maintaining data privacy. However, existing FL solutions are designed for static data and cannot support online training with continuous data streams from multiple sources. The main challenge lies in clients' (insurance companies) inability to automatically adapt to time-varying data flows in fraud detection tasks. To address this problem, we propose RegMeta-pFL, a personalized incremental FL scheme that combines Elastic Weight Consolidation (EWC) regularization and meta learning to effectively utilize incremental data from each client during training. Experiments on real-world insurance datasets show that RegMeta-pFL outperforms existing baselines, including FedAvg, FedProx, Per-FedAvg, and Fed_CIL, improving accuracy by 21.8%, 8.77%, 5.82%, and 2.71 % respectively, demonstrating its potential for fraud detection in the vehicle insurance sector.

With fault propagation in both physical and cyber networks, composite cascading failures pose great challenges to Cyber-Physical Systems (CPS), such as smart grids. Due to the complex coupling between networks, and the distinct propagation rules of faults originated from different kinds of nodes, its time-varying impact on the CPS is yet to be uncovered, which is essential to resilient CPS design. To study this problem, we establish a generic composite cascading model that incorporates both load redistribution and cyber virus infections. By this model, we observe an interesting change in both short-term and longterm evolution behaviors compared to its simple counterpart. Moreover, we propose an iterative inference algorithm to analyze its transient impact. Given the snapshot of any initial condition, extensive simulations show that our algorithm predicts nodal failing probabilities and the expected surviving ratio of the system with a narrow error margin (5 % and 15 %, respectively) under various settings, outperforming the data-driven method based on GNN + RNN. With these information, effective prevention and control measures of composite cascading failures can be designed for CPS.

With the continuous rolling-out of edge computing, Federated Learning (FL) has become a promising solution for intelligent Internet-of-things (IoT). In addition to resource constraints, deploying FL schemes in IoT networks is greatly challenged byheterogeneityin multiple dimensions. While heterogeneity in data distribution and computation capability has been extensively studied, the impact of distinct, even hybrid training paradigms on FL performances remains largely unknown. To answer this open question in the IoT context, we propose aHybrid-Training Federated Learning(HT-FL) algorithm for the power-constrained IoT networks, incorporating both sequential and parallel training that naturally adapts to various sub-network topologies, while greatly reducing the energy consumption during the training stage. We demonstrate through analysis that the convergence of HT-FL is theoretically guaranteed, achieving$O (\frac{1}{\sqrt{K}})$for carefully chosen learning rates. Experiments on multiple datasets show that, the proposed HT-FL outperforms existing FL schemes on multiple training tasks under various data distribution settings, while reducing an average of 20% energy consumption. In a more practical sense, a self-adaptive parameter-tuning strategy is also designed for HT-FL deployment, which can be easily extended to other multi-layer FL schemes in complex application scenarios.

The prevalence of location-based services has made positional information indispensable to everyday life, which has sparked growing concerns about the security of location privacy. In this paper, we propose to protect location privacy in wireless lo-calization from the perspective of power allocation. A closed-form expression of location secrecy metric (LSM) is first established to quantify the degree of risk that the location can be inferred by the eavesdropper. Then, a transmit power optimization problem constrained by the LSM lower bound is formulated. Through problem transformation and fractional programming, the optimal power allocation is ultimately obtained. Simulation results show that compared with the other existing strategies, the proposed power allocation strategy can effectively protect location privacy at the cost of smaller nosltioning accuracy loss.

Task offloading, which refers to processing (computation-intensive) data at facilitating servers, is an exemplary service that greatly benefits from the fog computing paradigm, which brings computation resources to the edge network for reduced application latency. However, the resource-consuming nature of task execution, as well as the sheer scale of IoT systems, raises an open and challenging question: whether fog is a remedy or a resource drain, considering frequent and massive offloading operations? This question is nontrivial, because participants of offloading processes, i.e., fog nodes, may have diversified technical specifications, while task generators, i.e., task nodes, may employ a variety of criteria to select offloading targets, resulting in an unmanageable space for performance evaluation. To overcome these challenges of heterogeneity, we propose a gravity model that characterizes offloading criteria with various gravity functions, in which individual/system resource consumption can be examined by the device/network effort metrics, respectively. Simulation results show that the proposed gravity model can flexibly describe different offloading schemes in terms of application and node-level behavior. We find that the expected lifetime and device effort of individual tasks decrease as$O({}{1}/{N})$over the network size$N$, while the network effort decreases much slower, even remain$O(1)$when load balancing measures are employed, indicating a possible resource drain in the edge network.

Cloudlets, which refer to the edge computing services deployed at the proximity of end devices, are key providers of connectivity, storage, and computation resources to many applications. While access to cloudlets is pervasive in typical settings, it can be difficult in challenging, even hostile environments, such as military or post-disaster scenarios, featuring multi-hop communication and energy-constrained end devices. In these cases, cloudlets may have become the only equipment powerful enough to execute life-critical applications, such as battle-field situation awareness, tactic cooperation, and search-and-rescue missions. Quality of these services is greatly influenced by the minimum time that a packet can be delivered, i.e., the cloudlet access delay (CAD), whose characteristics remain unknown. To address the open question of fast and efficient cloudlet access, we establish a packet mobility model that allows CAD and energy consumption to be analyzed as a function of the initial device-cloudlet distance. We find that the expected CAD scales either linearly or quadratically under distinct types of packet mobility, and the successful access rate (SAR) can be bounded by functions of the delay constraint. Based on these findings, we develop a packet shedding algorithm that saves 24% transmission power, and reduces the average CAD by 2%, while maintaining a similar SAR in simulated cloudlet access environments.

Spectrum activity surveillance (SAS) is essential to dynamic spectrum access (DSA)-enabled systems with a two-fold impact: it is a primitive mechanism to collect usage data for spectrum efficiency improvement; it is also a prime widget to collect misuse forensics of unauthorized or malicious users. While realizing SAS for DSA-enabled systems appears to be intuitive and trivial, it is, however, a challenging yet open problem. On one hand, a large-scale SAS function is costly to implement in practice; on the other hand, it is not clear how to characterize the efficacy and performance of monitor deployment strategies. To address such challenges, we introduce a three-factor space, composed ofspectrum,time, andgeographic region, over which the SAS problem is formulated by a two-step solution: 3D-tessellation for sweep (monitoring)coverageand graph walk for detectingspectrum culprits, that is, devices responsible for unauthorized spectrum occupancy. In particular, our system model transforms SAS from a globally collective activity to localized actions, and strategy objectives from qualitative attributes to quantitative measures. With this model, we design low-cost deterministic strategies for dedicated monitors, which outperform strategies found by genetic algorithms, and performance-guaranteed random strategies for crowd-source monitors, which can detect adversarial spectrum culprits in bounded time.

Emerging mobile applications enable people to connect with one another more easily than ever, which causes networked systems, e.g., online social networks (OSN) and Internet-of-Things (IoT), to grow rapidly in size, and become more complex in structure. In these systems, different, even conflicting information, e.g., rumor v.s. truth, and malware v.s. security patches, can compete with each other during their propagation over individual connections. For such information pairs, in which a desired information kills its undesired counterpart on contact, an interesting yet challenging question is when and how fast the undesired information dies out. To answer this question, we propose a Susceptible-Infectious-Cured (SIC) propagation model, which captures short-term competitions between the two pieces of information, and define extinction time and half-life time, as two pivots in time, to quantify the dying speed of the undesired information. Our analysis revealed the impact of network topology and initial conditions on the lifetime of the undesired information. In particular, we find that, the Cheeger constant that measures the edge expansion property of a network steers the scaling law of the lifetime with respect to the network size, and the vertex eccentricities that are easier to compute provide accurate estimation of the lifetime. Our analysis also sheds light on where to inject the desired information, such that its undesired counterpart can be eliminated faster.

Most existing studies for information dissemination in the online social network are based on variations of the classical epidemic model. In such a model, nodes recursively infect, or share information to, their neighboring nodes with a certain probability. The higher degree a node has, the more likely it gets infected by its neighbors. Although widely accepted, we found there are certain discrepancies between existing epidemic models and social interactions in reality. Firstly, the real-world social network is actually a dual-layered network, where a person shares information online to her online friends, and also offline to her real-life friends. More importantly, since a computer do not automatically share information, a computer exposed to information will not effectively receive it (i.e., getting infected and starting to infect others) unless its user receives it. Secondly, contrary to the epidemic model, the more friends a person has, the less likely she is going to effectively receive a certain piece of message (just imagine how easily a message can be flushed and ignored by a human user because of overwhelming newer information). In other words, in social networks, the infection rate of a node may not be positively correlated with its degree. Based on these observations, we develop the social-physical interdependent (SPI) model to capture and analyze the unique characters of social networks. Our study provides new observations, and sheds light on a new direction for the study of information dissemination in social networks.

Spectrum monitoring, run-time usage acquisition, and regulation enforcement, in general can be referred to as spectrum activity surveillance (SAS). It is essential to dynamic spectrum access with a two-fold impact: it is a primitive mechanism to continuously scan spectrum usage for system optimization purposes; it is also a prime widget to obtain spectrum footprints of legitimate users, and record misuse by unauthorized or malicious users. Seemingly trivial, large-scale SAS in wireless overlay networks is actually an open yet challenging problem. This is because on one hand, such a system is time and energy-sensitive and hence unlikely (or not necessary) to implement in practice, due to constraints of radio spectrum license and system deployment. On the other hand, it is not clear how to characterize the efficacy and performance of spectrum monitoring strategies in surveillance over a large geographical region, and detection of spectrum culprits, that is, unauthorized spectrum occupants. To address such a challenge, we consider SAS in a 3-dimensional space that is composed of spectrum, time, and geographical region, and then formulate monitoring strategies as graph walks by accounting for the locality of spectrum activities. In particular, our approach transforms the SAS problem from a globally collective activity to a set of localized, distributed actions, and strategy objectives from qualitative attributes to quantitative measures. We find that randomized strategies with m monitors can achieve a sweep-coverage over a space of n assignment points in Θ(n/m ln n) time, and detect an oblivious or adversarial spectrum culprit in Θ(n/m) time for SAS systems.

Internet of Things (IoT) is a networking paradigm that interconnects physical systems to the cyber world, to provide automation and intelligence via interdependent links between the two domains. Such interdependence renders IoT systems vulnerable to random failures, e.g., broken communication links or crashed cyber instances, because a single incident in one domain can develop into a cascade-of-failures across domains, which dissolves the network structure, and has devastating consequences. To answer how robust an IoT system is, this paper studies its resilience by examining the impact of edge- and jointly-induced cascades, that is, a sequence of failures caused by randomly broken physical links (and simultaneous failing cyber nodes). Resilience of an IoT system is quantified by two new metrics, the critical edge disconnecting probability φcr, i.e., the maximum intensity of random failures the system can withstand, and the cascade length τcf, i.e., the lifetime of a cascade. For IoT systems with Poisson degree distributions, we derive exact solutions for the critical disconnecting probability φcr, above which an edge-induced cascade will completely fragment the network. We also find that the critical condition φcrmarks a dichotomy of the expected cascade length E(τcf): for the super-critical (φ > φcr) scenario, we obtain E(τcf) ~ exp(1 - φ) through analysis, while for the subcritical scenario, we observe E(τcf) ~ exp(1/1 - φ) through simulations. With these results, the final outcome of a cascade can be anticipated upon the initial failures, while the reaction window of time-sensitive countermeasures can be obtained before a cascade fully unfolds.

Internet of things (IoT) is expected to provide a fully informative and controllable environment that features networking, automation, and intelligence by interconnecting physical systems to cyber world. Such a correlation opens the interdependence between the two, upon which a single incident in one domain, e.g., a broken communication link, or an out-of-battery device, can cause a cascade-of-failures across physical and cyber domains. To understand the resilience of IoT systems against such detrimental cascades, this paper studies the aftermath of edge and jointly-induced cascades, that is, a sequence of failures induced by randomly broken physical links (and simultaneous failing cyber nodes) by answering how many nodes will survive the cascade with a newly defined node yield metric. Specifically, we construct a framework to establish self-consistent equations of node yield through an auxiliary graph, without requiring the exact network topology. Then two algorithms are proposed to numerically calculate node yield for interdependent networks with arbitrary degree distributions. For random graph with Poisson degree distributions, we prove the existence of a critical initial edge disconnecting probability φcr, under which an edge-induced cascade will result in dissolving the network topology, derive the closed form solution for φcr, and find that φcrincreases sub-linearly with the mean degree of the physical network.

Determining the causality of abnormalities in a network is the prerequisite for developing countermeasures. In this paper, we focus on infection detection in heterogeneous networks. Given a snapshot of the network which demonstrates the condition of the nodes, the goal is to distinguish between random failures and epidemic scenarios. We model the network situation as a graph signal based on the nodes' status. Detection metrics motivated by graph signal processing are introduced for the infection detection problem in hand, and an effective algorithm is proposed to solve it. Simulation results indicate a dramatic improvement in terms of detection probability compared to the current state-of-the-art.

Spectrum monitoring is a prerequisite in dynamic access regulation, policy enforcement, as well as spectrum database establishment. In this paper, we introduce the dimension of geographical space into the spectrum monitoring problem, and studied deployment strategies of multiple monitors, in terms of coverage time and cost. The monitoring problem is modeled as a 3-d continuous sweep coverage problem, whose solution space is then reduced by effectively dividing the spectra-location space, in order to achieve a small coverage time. The cost minimization is then formulated as a Multiple Traveling Salesman problem (MTSP), which is NP-hard. By observing the structure of the strategy space, we propose a solution that attains a reasonable cost, without applying complex optimization algorithms.

As online social networks grow in both size and connectivity, epidemic information dynamics in such networks is attracting considerable research interests, due to its impact on both the network and individuals. This paper studies control of malicious information (virus) epidemic with replicable antidote information, taking topological characteristics of the underlying graph into consideration. Specifically, we analytically relate the extinction time of the virus to the diameter and giant component size of the remaining graph after the initial antidote distribution. With this divide and conquer guideline, topology-based antidote distribution approaches are designed, and then examined through simulations in real world network portions.

Online Social Networks (OSNs) is taking over television and newspapers, to be the dominant information dissemination option. The growing involvement of individuals create the situation that colliding, even contradicting information coexist and propagate in the same network, which gives rise to an interesting question: how will the conflicting information propagate? To answer this question, the propagation process is described to be an Susceptible-Infected-Cured (SIC) epidemic, and we propose an inference algorithm to study the transient behavior of the competing propagation processes in connected networks. Moreover, we provide an analytic method to derive the conditional infection count distribution for networks with special topologies, as a step further to understand the evolution. A trace collected from the Internet is analyzed to validate our model and methods.

In an era of networks in which any individual is connected with one another, such as Internet of Things (IoT) and Online Social Networks (OSNs), the networks are evolving into complex systems, carrying a huge volume of information that may provoke even more. An interesting, yet challenging question is how such information dissemination evolves, that is, to continue or to stop. Specifically, we aim to find out the aftermath of epidemic spreading via individuals and conflicting information dissemination. From a holistic, networking view, it is impossible to take every aspect into accounts for complex networks toward these questions. Therefore, we establish a Susceptible-Infectious-Cured (SIC) propagation model to examine two simple network topologies, clique and star, in terms of extinction time and half-life time of information under controllable, epidemic dynamics. For a network of size n, both theoretical and numerical results suggest that extinction time and half-life time are O(log n/n) for clique networks, and O(log n) for star networks. More interestingly, given an initial network state I0, the extinction time is constant (O(1)) for cliques, and O(log I0) for stars; while the half-life time is O(log 1/I0) for both clique and star networks, respectively. In addition, we developed a method to estimate the conditional infection count distribution, which indicates the scope of information dissemination.