Edge computing (EC) consists of deploying computation resources close to the users, thus enabling low-latency applications, such as augmented reality and online gaming. However, large-scale deployment of edge nodes can be highly impractical and expensive. Besides EC, there is a rising concept known as Vehicular Cloud Computing (VCC). VCC is a computing paradigm that amplifies the capabilities of vehicles by exploiting part of their computational resources, enabling them to participate in services similar to those provided by the EC. The advantage of VCC is that it can opportunistically exploit part of the computation resources already present on vehicles, thus relieving a network operator from the deployment and maintenance cost of EC nodes. However, it is still unknown under which circumstances VCC can enable low-latency applications without EC. In this work, we show that VCC has the potential to effectively supplant EC in urban areas, especially given the higher density of vehicles in such environments. The goal of this paper is to analyze, via simulation, the key parameters determining the conditions under which this substitution of EC by VCC is feasible. In addition, we provide a high level cost analysis to show that VCC is much less costly for a network operator than adopting EC.

This work was accepted to be published at IEEE WCNC'24.

Vehicular Ad hoc Networks (VANETs) offer a promising approach to enhancing road safety. Cooperative Awareness Messages (CAM) is an essential service in VANETs, allowing vehicles to transmit radio beacons containing their positions and velocities. These messages inform nearby vehicles about the traffic situation. This paper focuses on Extended Cooperative Awareness Messages (ECAM), which include additional information about nearby vehicles. ECAM beacons consist of a vehicle's speed, position, and data on the positions and velocities of other vehicles in its vicinity. This comprehensive information enables nearby vehicles to understand the traffic situation and take appropriate actions to prevent potential collisions. Studies demonstrate that ECAM has the potential to significantly improve road safety by providing comprehensive and up-to-date traffic information. This paper uses stochastic geometry to evaluate different versions of ECAM services and compare the results with simple simulations. The evaluation assumes random vehicle placement using a homogeneous Poisson Point Process and models the ECAM service using the Matern Point Process.

This article was published at PEMWN'23 25.

As 5G Cellular Vehicle-to-Everything (C-V2X) technology takes the lead in V2X communication, it opens the possibility for telecommunication service providers to offer Vehicleto-Network (V2N) services using their existing 5G network infrastructure. To enhance the security of 5G V2N services, in this paper we propose a novel collaborative V2X misbehavior detection system. This system would safeguard the V2X application servers (V2X ASs), deployed in the 5G edge network, from any malicious V2X position manipulation attacks. Our proposal includes two enhanced machine learning models. The first model utilizes historical data to conduct On-Road Plausibility Checks (ORPC), while the second model builds upon the first by enabling collaboration among edge detection nodes through the sharing of attack ratios for each vehicle. Our proposed models were tested using extensive 5G core-network emulations, yielding excellent results. The first model achieved a notable accuracy improvement from 73% to 91%, while the second model further enhanced the accuracy to an impressive 95%.

This article was presented at GLOBECOM'2023

The emergence of 5G Cellular Vehicle-to-Everything (C-V2X) has made it the predominant technology for enabling Vehicle-to-Everything (V2X) communications. As a result, this has created an opportu- nity for telecommunications service providers to leverage their pre-existing 5G network infrastructure, enabling them to provide Vehicle-to-Network (V2N) services. In this paper, we propose a new approach that enhances the security of 5G V2N services through the implementation of a Federated Learning V2X misbehavior de- tection system within the 5G core network. The proposed system aims to protect V2X application servers (V2X ASs) that are located in 5G edge networks against potential V2X attacks while leveraging the privacy and scalability advantages of Federated Learning. Our proposed model is compared, using extensive emulations, to other centralized and distributed approaches, achieving excellent results, which makes it feasible for deployment. Our proposal achieved a notable accuracy of 98.4%, while scoring an impressive 99.3% precision and 96.9% detection rate.

This article was presented at MSWiM'2023 and was the best paper runner-up.

One of the main difficulties to enable the future scaling of IoT networks is the issue of massive connectivity. Recently, Modern Random Access protocols have emerged as a promising solution to provide massive connections for IoT. One main protocol of this family is Irregular Repetition Slotted Aloha (IRSA), which can asymptotically reach the optimal throughput of 1 packet/slot. Despite this, the problem is not yet solved due to lower throughput in non-asymptotic cases with smaller frame sizes. In this paper, we propose a new variant of IRSA protocol named Deep-Learning and Sensing-based IRSA (DS-IRSA) to optimise the performance of IRSA in short frame IoTs, where a sensing phase is added before the transmission phase and users' actions in both phases are managed by a deep reinforcement learning (DRL) method. Our goal is to learn to interact and ultimately to learn a sensing protocol entirely through Deep Learning. In this way, active users can coordinate well with each other and the throughput of the whole system can be well improved. Simulation results show that our proposed scheme convergence quickly towards the optimal performance of almost 1 packet/slot for small frame sizes and with enough minislots and can achieve higher throughput in almost all cases.

This article 17 was presented at GLOBECOM 2003.

A multi-level random power transmit strategy that is used in conjunction with a random access protocol (RAP) (e.g. ALOHA) is proposed to fundamentally increase the throughput in a distributed communication network. A SIR model is considered, where a packet is decodable as long as its SIR is above a certain threshold. In a slot chosen for transmission by a RAP, a packet is transmitted with a power level chosen according to a distribution, such that multiple packets sent by different nodes can be decoded at the receiver in a single slot, by ensuring that their SIRs are above the threshold with successive interference cancelation. The achievable throughput and the upper bounds are shown to be close with the help of comprehensive simulations. The main takeaway is that the throughput of more than 1 is possible in a distributed network, by using a judicious choice of power level distribution in conjunction with a RAP.

This paper was presented at the conference NCC 2023, Guwahati, India.

Recently, two dimensional orthogonal time frequency space (OTFS) modulation technique has introduced in wireless communications to combat the effects of multipath fading and Doppler spread. In this paper, we analyze the impact of nonlinear power amplifier (NPA) and phase noise on the OTFS system over the EVA channel model. The study focuses on the bit error rate (BER) performance concerning the input-back-off (IBO) and the nonlinearity parameter values of the NPA. The results demonstrate that both NPA and phase noise significantly degrade the OTFS system performance, especially for higher modulation schemes and low values of IBO. Furthermore, we numerically analyze the impact of system parameter variations on BER performance.

This work was presented at IEEE International Conference on Advanced Networks and Telecommunications Systems ANTS 2023.

The study introduces active intelligent reflecting surfaces (AIRS) as a solution to improve signal strength in wireless systems, overcoming the limitations of passive IRS (PIRS). AIRS amplifies and adjusts the phase of reflected signals, enhancing system performance. The research focuses on analyzing the bit error rate and sum-rate of an AIRS-assisted system, considering various factors like channel information and distance between components. Simulations reveal that AIRS significantly outperforms PIRS, especially in scenarios with a direct link. However, a highly equipped PIRS can approach AIRS's efficiency under certain conditions.

This work was presented at IEEE International Conference on Advanced Networks and Telecommunications Systems ANTS 2023.

As the main goal of connected autonomous vehicles' communications is to improve traffic safety and save lives, any design of a resource allocation scheme must consider the stringent requirements of these applications in terms of latency and reliability. For this, 5G cellular networks suitably address these challenges. This paper proposes a new mechanism for the telco operator to adapt the physical (PHY) layer configuration for efficient radio resource management in a 5G New Radio (NR) based system. To tackle this issue, we propose to adjust the PHY layer numerology configuration by fine-tuning it with a Radio-Aware Adaptive PHY Layer Configuration (RA-APC) algorithm in order to maximize the Effectively Transmitted Packet (ETP). Extensive simulations show that our proposal RA-APC achieves strong improvements in terms of ETP, reliability and latency while considering safety and non-safety traffic scenarios.

This article was presented at ICC 2023.

5G Cellular Vehicle-to-Everything (C-V2X) is expected to become the dominant technology to enable Cooperative Intelligent Transport System (C-ITS) applications. In this paper we address the problem of detecting falsified vehicle positions sent by misbehaving vehicles targeting C-ITS application servers over 5G networks. We propose a novel security system as 5G application function. It is based on machine learning and integrated with the 5G core network to monitor, detect and prevent potential misbehavior. Based on extensive network simulations utilizing 5G network emulator, our proposal achieves very good performances, accurately reported 98

This article was presented at ICC 2023.

Our work aims to design protocol sequences through deep reinforcement learning (DRL). Protocol sequences are periodic binary sequences that define multiple access control among users, introduced for systems considering collision channel without feedback (CCw/oFB). In this paper, we leverage the recent advancement of DRL methods to design protocol sequences with desirable new properties, namely Throughput Maximizing User- Irrepressible (TMUI) sequences. TMUI has two specific properties: (i) user-irrepressibility (UI), and (ii) maximizing the minimum individual throughput among the users. We assumed that the transmission channel is divided into time slots and the starting time of each user in joining the system is arbitrary such that there exist random relative time offsets. We use a DRL approach to find TMUI sequences. We report the obtained TMUI protocol sequences and conduct numerical studies comparing TMUI against slotted ALOHA. Simulation results also show that the new medium access control (MAC) protocol does hold the UI property and can achieve substantially higher minimum individual user throughput, under the same system parameters.

The article 15 was presented at EuCNC & 6G Summit 2023 - European Conference on Networks and Communications & 6G SummitGothenburg, SwedenJun 6, 2023.

This work exploits the potential of two important technologies which are UAVs (Unmanned Aerial Vehicles) and blockchain in the context of Agriculture 4.0. We propose a cattle health monitoring system based on UAVs that collect health measures from IoT devices equipping the animals. The main objectives of our system are twofold. First, the consumer will be aware of the quality of his/her food. Second, the national ecosystem (e.g. agriculture ministry, trade ministry) will get useful information about the quality and the number of cattle that can be put on the market. Thus, smart cattle management strategies could be undertaken afterward. The involved entities in such a system are multiple: the farmer, the veterinaries, the Ministry of Agriculture, etc... We first start by studying the system's security by applying the FMEA risk assessment methodology. Our findings motivate us to integrate blockchain technology to manage the data collected as well as the attribution of the UAVs missions via a marketplace. Thanks to its properties, this technology ensures the transparent tracking of cattle status and fairness in the payment of the UAVs managed by private operators. Finally, we develop a proof of concept using the Sui blockchain platform.

This work was published in the conference International Wireless Communications and Mobile Computing, IWCMC 2023.

Multi-hop wireless networks can be optimized using directional antennas, particularly in scenarios like drone networks, where link performance is heavily dependent on nodes' positions. This optimization ensures high operational guarantees, instantaneous connectivity, minimum SNRs and SINRs thresholds, and improved QoS. It simplifies tasks of future network layers and allows for more relaxed routing protocols and scheduling. However, attaining optimal performance via network configuration, which involves selecting an antenna orientation for each node to create a link with another node, is challenging, especially when it is carried out online and in real-time. This task is highly combinatorial and is often treated as an Integer Programming (IP) problem. While it is suitable for static graphs or situations where real time solution computing is not needed, when dealing with dynamic networks that need frequent and important topology changes, one may need a lighter, offline solution.

To tackle this challenge, we present nodes2net, a method grounded in deep learning for configuring network topology. This approach uses nodes' positions as inputs and produces a set of links as output. We teach it to imitate ideal graphs obtained by more costly algorithms such as IP methods. By leveraging learning of patterns and theoretically driven properties, nodes2net can generate reliable network configuration solutions when dealing with new sets of node positions. It utilizes efficient neural network aggregation operators to facilitate and process information about the nodes, to finally produce the final solution as set of links. Our results demonstrate the competitive performance of this method.

This article was presented at PEMWN'2023.

It is common knowledge that using directional antennas is often mandatory for Multi-hop ad-hoc wireless networks to provide satis- fying quality of service, especially when dealing with an important num- ber of communication nodes [1]. As opposed to their omnidirectional counterpart, directional antennas allow for much more manageable in- terference patterns: a receiving antenna is not necessarily interfered by nearby emitting antennas as long as this receiving antenna is not di- rected towards these undesired emission beams. Two nodes then need to steer one of their antennas in the direction of the other node in order to create a network communication link. These two users will then be able to, in turn, emit and receive to and from each other. The scope of this work resides in finding a centralized algorithm to governate these antenna steering decisions for all users to instantaneously provide a valid set of communication links at any time given the positions of each user. The problem that raises is then a geometrical one that implies finding topologies of network links that present satisfying throughput and over- all QoS and guarantee instantaneous connectedness i.e. the computed set of links allows any user to reach any other user in a certain number of hops. Building such optimized link topologies makes further tasks, such as routing and scheduling of the network, much simpler and faster. This problem is highly combinatorial and, while it is solvable with traditional Mixed Integer Programming (MIP), it is quite challenging to carry it out in real time. For this purpose, we propose a Deep Neural Network that is trained to imitate valid, solved instances of the problem. We use the Attention mechanism to let nodes exchange information in order to capture interesting patterns and properties that then enable the neu- ral network to generate valid network link topologies, even dealing with unseen sets of users positions

This article was presented at MLN'2023.

We have investigated mobile networks of Unmanned Aerial Vehicles (UAVs) to extend connectivity and guarantee data rates in the 5G by analyzing possible hovering locations based on limitations such as flight time and coverage. We provide analytic bounds on the requirements in terms of connectivity extension for vehicular networks served by fixed Enhanced Mobile BroadBand (eMBB) infrastructure, where both vehicular networks and infrastructures are modeled using stochastic and fractal geometry as a model for urban environments.

We prove that assuming $n$ mobile nodes (distributed according to a hyperfractal distribution of dimension $d_F$) and an average of $\rho$ Next Generation NodeB (gNBs), distributed like a hyperfractal of dimension $d_r$ if $\rho =n^{\theta }$ with $\theta >d_r/4$ and letting $n$ tending to infinity (to reflect megalopolis cities), then the average fraction of mobile nodes not covered by a gNB tends to zero like $O\left(n^{-\frac{(d_F-2)}{d_r}(2\theta -\frac{d_r}{2})}\right)$. Interestingly, we prove that the average number of drones, needed to connect each mobile node not covered by gNBs is comparable to the number of isolated mobile nodes. We complete the characterization by proving that when $\theta <d_r/4$ the proportion of covered mobile nodes tends to zero. The hyperfractal model can be used to model cities with very few parameters. Furthermore it can be run as a generative models to create an unbounded number of imaginary cities for AI training.

This work has been published in 49.

PhilippeJacquet

Blockchain applications continue to grow in popularity, but their energy costs are clearly becoming unsustainable. In most cases, the primary cost comes from the amount of energy required for proof-of-work (PoW). Here we study the application of blockchains to the IoT, where many devices are underpowered and would not support the energy cost of proof of work. PoW was originally intended for two main uses: block moderation and protecting the blockchain from tampering. The first use is by far the most energy eater. It has already been proposed to replace the expensive moderation by PoW with the energy-efficient green mining. Free from the block moderation burden, the PoW can be made much lighter and adapted to the power diversity of the miners. We propose a fractal difficulty PoW. Used alone we show that the fractal PoW does not really reduce the energy cost for the low powered nodes. However when associated with green election which guarantees a finite period of fairness indifferent to PoW after each block mined, we show that the fractal PoW indeed reduces the energy for the low powered devices while keeping the same protection against blockchain tempering. In passing we show that a certain monotonicity condition is not met by PoW.

This work has been presented in PEMWN 2023 28

In this work, we aim at answering the question “at what frequency should one sample individual human movements so that they can be reconstructed from the collected samples with minimum loss of information?”.

Our analyses on fine-grained GPS trajectories from users around the world unveil (i) seemingly universal spectral properties of human mobility, and (ii) a linear scaling law of the localization error with respect to the sampling interval. Our analysis of fine-grained trajectories unveils a novel linear scaling law of the localization error with respect to the sampling interval. Such results were published at IEEE Globecom 2017  40.

Building on these results, we challenge the idea of a fixed sampling frequency and present a lightweight, energy-efficient, mobility aware adaptive location sampling mechanism. We thus present DUCTILOC, a location sampling mechanism that takes advantage of the law above to profile users and then adapt the position tracking frequency to their mobility. Our design is energy efficient, as DUCTILOC does not rely on power-hungry sensors or expensive computations; moreover, it provides a handy knob to control energy usage, by configuring the target positioning accuracy. Real-world experiments with an Android implementation of DUCTILOC show that it can effectively adjust the sampling frequency to the mobility habits of each individual and target accuracy level, reducing the energy consumption by 60% to 98% with respect to a baseline periodic sampling.

This work is published at the IEEE ACCESS journal 14.

WiFi-based crowdsensing is a major source of data in a variety of domains such as human-mobility, pollution-level estimation, and, opportunistic networks. MAC randomisation is a backbone for preserving user-privacy in WiFi, as devices change their identifiers (MAC addresses). MAC association frameworks in the literature are able to associate randomized MAC addresses with a device. Such frameworks facilitate the continuation and validity of works based on device-based identifiers. In this paper, we first question and verify the reliability of these frameworks with respect to the datasets (scenarios) used for their validation. Indeed, we observe a substantial discrepancy between the performances obtained by these frameworks when confronting them with different contextual environments. We identify that the device heterogeneity in the input scenario is privacy-preserving. Henceforth, we propose a novel metric: randomization complexity, capable of successfully catching the degree of randomization in evaluated datasets. Existing and new frameworks can thus be benchmarked using this metric to ensure their reliability for any datasets with similar or lower randomization complexities. Finally, we open discussions on the potential impact of the benchmarks in the domain of MAC randomization.

This article was published at IEEE VTC'23 23 and is part of the contributions described in Abhishek's PhD thesis 32. He did a PhD under the supervision of Aline C. Viana and Nadjib Achir. He defended in October 2023.

Smartphones or similar WiFi-enabled devices regularly discover nearby access points by broadcasting management frames known as probe-requests. Probe-request frames relay, as information, the MAC addresses of sending devices, which act as the device identifiers. To protect the user's privacy and location, probe-requests use a randomized MAC address generated according to the MAC address randomization protocol. Unfortunately, MAC randomization greatly limits any studies on trajectory inference, flow estimation, crowd counting, etc. To overcome this limitation while respecting users' privacy, we propose Bleach, a novel, efficient, and comprehensive approach allowing randomized MAC addresses to device association from probe-requests. Bleach models the frame association as a resolution of MAC conflicts in small time intervals. We use time and frame content-based signatures to resolve and associate MACs inside a conflict. We propose a novel MAC association algorithm involving logistic regression using signatures and our introduced time metric. To the best of our knowledge, this is the first work that formulates the probe-request association problem as a generic resolution of conflicts and benchmarks the association with respect to several datasets. Our results show that Bleach outperforms the state-of-the-art schemes in terms of accuracy (as high as $99\%$) and robustness to a wide range of input probe-request datasets.

This work is related to the ANR MITIK project (2020-2025). It is under-submission to a journal and is described in Abhishek's PhD thesis 32, a PhD performed under the supervision of Aline C. Viana and Nadjib Achir. He defended in October 2023.

A preliminary work discussing privacy flaws in WiFi standards was published at the IEEE LCN 2021 (Doctoral-track - Promising ideas)  45 and motivated the here above MAC association work.

Human mobility is challenging to infer, reconstruct, or predict precisely and even further through a more privacy-preserving and scalable manner. Domains and applications are: targeted advertising, epidemic prevention, urban, transportation, or touristic planning, to cite a few. Current GPS-based localization methods are considered sparse in space and time, and RSSI-based passive sniffing methods are challenging due to miscellaneous error sources. Recent literature has shown the large and highly volatile errors in human-location estimation when using observed RSSI from passive sniffing over Wireless packets.

In this work, we propose the first framework that introduces the concept of the user's bounded trajectory. We propose to leverage the signal strength of users' public WiFi probe requests collected from measurements of multiple deployed WiFi sniffers. First, we investigate and characterize errors in RSSI-based radial distance (between the user and each sniffer) estimation. Then, we approximate such radial distances leverage and deduce bounds associated with a user's position. Finally, we infer a user's bounded trajectory using the spatiotemporal bounds of users' locations over time. We guarantee the bounds to enclose a user in space and time, with $95\%$ confidence and a $10\%$ margin of error. Using real-world and large-scale synthetic datasets under heterogeneous contexts and wireless conditions, we infer trajectories with bounds' width of less than $10m$ in $70\%$ of cases with users' inclusiveness close to $100\%$.

This work is related to the ANR MITIK project (2020-2025). It was published at WCNC 2023 21 and is also described in Abhishek's PhD thesis 32. He did a PhD under the supervision of Aline C. Viana and Nadjib Achir. He defended in October 2023.

The increasing proliferation of Wireless Fidelity (WiFi) and Bluetooth Low Energy (BLE) networked devices broadcasting over-the-air unencrypted public packets has raised growing concerns regarding users' privacy. Such public packets consist of management frames, like probe-requests and beacons, necessary for devices to discover available wireless networks and enhance user experience. Revealing the MAC address of a device through public packets allows adversaries to follow the device and do behavioral profiling. Modern devices periodically change/randomize their advertised MAC addresses. Nevertheless, attacks on MAC address randomization have been carried out, demonstrating that randomized addresses from a device can be associated with as little information as the timestamps of their advertised public packets.

In this thesis, we identify key flaws that lead to the MAC association. To measure the severity of identified flaws by looking at the performance of current MAC association attacks, we need large-scale traces of public packets with "ground truth" information regarding randomized addresses from the same device. We assert the flaws by employing our proposed simulation framework to generate large-scale WiFi and BLE passive sniffing traces. We reveal that current device randomization is ineffective and needs revision.

In addition to key flaws identifications, we address the unreliability of existing association frameworks with respective trace collection scenarios to understand the factors contributing to variable association performance. We conduct case studies and introduce benchmarks for evaluating the performance of any association framework. We show the need for a new and effective WiFi MAC association framework, and finally, we develop and benchmark a novel association framework to determine its expected performance with any new input probe-request dataset.

Once achieving effective MAC association, we reveal the inference of user locations from passively sniffed probe-requests. In this thesis, we identify the limitations of the Received Signal Strength Indicator (RSSI) in accurately inferring user trajectories as a series of timestamped locations due to its high variability. Considering this, we propose a novel concept called "bounded trajectories." Bounded trajectory refers to an area where a particular user is probable to be present across time. We analyze and model the errors associated with radial-distance estimation to derive bounded trajectories that offer high inclusiveness of users' actual trajectory and narrow width throughout its course.

This PhD thesis 32 was performed by AbhishekKumar Mishra under the supervision of Aline C. Viana and Nadjib Achir and under the ANR MITIK project's funding. Abhishek defended his phd on the 19th October 2023 and is currently a Post-Doc fellow at the PRIVATICS team of INRIA Lyon.

The study of human mobility is fundamental because of its impact on urban planning, the spread of diseases, the well-being of the population, and the mitigation of pollution, among other applications. Among the open challenges in the area, we have as one of the most important the interpretability and generality of the generated models, and the unbalanced volume of available data; several areas have little data available, making it impossible to use existing models.

We intend to face these challenges in a way not addressed before in the literature, which is with the use of mathematical models inspired by natural phenomena—normally modeled as differential equations—combined with established ML techniques to develop prediction models in the area of mobility. We intend with this combination to bring more interpretability to the models and reduce the need for large volumes of data.

This work is focused on the area of aggregate mobility prediction because of the data that we have to carry out this work. The available data describe the flow of people between administrative regions of Paris, France, with a sampling frequency of one hour and during fourteen days. More specifically, we intend to model the visitation routine of the people to predict the population density of areas in an instant of time, thus considering mobility and people's routine as a phenomenon to be modeled.

Our work mainly seeks to answer the following problem: Is it possible, with a good level of assertiveness, to model people's visitation routine through mathematical models combined with machine learning? Several works available in the literature show that the movements of people are typically characterized by routine behavior: daily cyclical movements (home to work), few places visited in their routine, and displacements that reveal preferred trajectories. The use of mathematical models to add domain knowledge of mobility as a phenomenon in ML techniques is new in the area and to advance this study, bringing more applicability to models, is a valuable knowledge gain for applications in urban planning or epidemiology.

We believe that (in addition to producing a generic interpretable model and requiring less training data to predict the number of people present in each of the study regions at an instant of time) our thesis will open up new opportunities for the development of mobility prediction models that consider other aspects, such as the trajectories that are expected to be taken by individuals or groups of individuals.

Haron has defended his PhD mid-term examen and the work is still on-going.

SafeCitymap is a data-driven project investigating how individuals' mobility patterns at a metropolitan scale were affected by the Covid-19 pandemic, and especially by the harsh French lockdown conditions enforced from March 17, 2020 to May 11, 2020 (i.e., two weeks before and during the first, second, and third lockdown). For this, we used spatiotemporal aggregated mobile phone data provided by SFR, a major SFR French telecom operator, covering a geographical region focused on the city of Paris. An essential property of this data is its fine-grained spatial resolution, which, to the best of our knowledge, is unique in the COVID-related mobility literature.

We perform a data-driven mobility investigation and modeling to quantify (in space and time) the population attendance and visiting flows in different urban areas. Second, when looking at periods both before and during the lockdown periods, we quantify the consequences of mobility restrictions and decisions on an urban scale. For this, per zone, we define a so-called signature, which captures behaviors in terms of population attendance in the corresponding geographical region (i.e., their land use) and allows us automatically detect activity, residential, and outlier areas. We then study three different types of graph centrality, quantifying the importance of each zone in a time-dependent weighted graph according to the habits in the mobility of the population. Combining the three centrality measures, we compute per zone of the city, its impact-factor, and employ it to quantify the global importance of zones according to the population mobility.

Our results firstly reveal the population's daily zone preferences in terms of attendance and mobility, with a high concentration on business and touristic zones. Second, results show that the lockdown mobility restrictions significantly reduced visitation and attendance patterns on zones, mainly in central Paris, and considerably changed the mobility habits of the population. As a side effect, most zones identified as mainly having activity-related population attendance in typical periods became residential-related zones during the lockdown, turning the entire city into a residential-like area. Shorter distance displacement restrictions imposed by the lockdown increased visitation to more “local” zones, i.e., close to the population's primary residence. Decentralization was also favored by the paths preferences of the still-moving population. On the other side, “jogging activities” allowing people to be outside their residences impacted parks visitation, increasing their visitation during the lockdown. By combining the impact factor and the signatures of the zones, we notice that areas with a higher impact factor are more likely to maintain regular land use during the lockdown.

Results of such investigations are periodically posted in the interactive webpage here: SafeCityMap website. This work is published at the ACM Transactions on Spatial Algorithms and Systems  38 and a extended report is available at  37.

Currently, we are investigating if the previously described mobility modeling can be used as proxies for the inference of pollution and noise indicators in a metropolitan city. While the polution and noise depend on physical deployed sensors, mobility datasets provide more precise spatiotemporal information of larger geographical areas. Preliminary investigations show a strong correlation between such indicators and SafeCityMap mobility modeling. A new report describing such investigations is being prepared.

The Federated Learning (FL) framework has been applied in multiple domains, offering solutions that provide both accuracy and data privacy protection. FL has been applied to various problems, from image classification to next word prediction. For mobility prediction specifically, some prior efforts adapted solutions from other domains to mobility problems: e.g., using image classificaiton model for transportation mode prediction by converting coordinates into pixels). Yet, directly adapting solutions from other domains can be both challenging and inefficient due to the spatial and temporal related specificities of mobility prediction. Indeed, visits to certain point of interests are directly correlated to the users' routines and preferences. These are hard to embed in models for next word prediction or image classification, which are more concerned with grammatical structures of a language and recognition of patterns on a static low-dimensional space.

Yet, specifically for human mobility prediction, prior solutions have neglected the impact that the naturally heterogeneous human patterns may have on FL effectiveness. Also, such prior solutions are based on the social network datasets, that are both sparse in space and coarse in time, challenging routine and mobility patterns characterization. Hence, prior evaluations of FL on mobility prediction are limited and may overestimate the robustness of the proposed solutions.

In this work, we aim to fill this gap by analyzing the performance of alternative FL-based mobility predictions in scenarios with users with varying mobility patterns, identified in real and less sparse human mobility data. We aim to answer the following question: How do existing FL-based mobility prediction models perform for users with very different mobility patterns, such as very repetitive behavior (e.g., routines) or more exploratory visiting patterns (e.g., tourists)? Our analyses comprise both model effectiveness (accuracy) and efficiency (resource usage and execution time), and offer insights into possible improvements to current FL solutions. Heterogeneity on the (fine-grained) spatial and temporal mobility patterns directly impact prediction, hardening the FL performance analysis.

How preliminary results show considerable impact that different mobility patterns can have on both the effectiveness and efficiency of the FL models. Easier to predict users (i.e., with a strong routini in their daily life: visite regular locations and are very often stationary) experience great improvements on the accuracy of both models, accelerated the learning process and reduced resource consumption. Contrarily, users with less routine, i.e., that are more prone to explore and discover new areas, have challenged both models on every aspect, specially accuracy. Indeed, even a small fraction of them (14%) greatly impacted FL models in some senarios, evidencing the impact that heterogeneity has on the solutions.

These initial results offer many future directions of exploration. For example, we aim to study strategies to explicitly incorporate into model training the different properties of the mobility profiles, as well as investigate the performance trade-offs of favoring one particular profile over the others. Studying how the FL solutions can adapt to fluctuations in mobility patterns (and even profiles) over time is also worth pursuing.

A preliminary version of this work was published as a poster at the NetMob 2023 conference. This work is on-going and constitutes the Master thesis of Joao Esper to be submissed on Mars 2024.

Several studies on the analysis of human mobility patterns have been carried out focusing on the identification and characterization of important locations in users' life in general. We extended these works by studying human mobility from the perspective of mobile data offloading. In our first study, We define Offloading Regions (ORs) as areas where a commuter's mobility would enable offloading, and propose an unsupervised learning method to extract ORs from mobility traces.

Next, we leverage human mobility to inform offloading tasks, taking a data based approach leveraging granular mobility datasets from two cities: Porto and Beijing. We evaluate the offloading opportunities (ORs) provided to users while they are travelling in terms of availability, time window to offload, and offloading delay. Results show that in 50% of the trips, users spend more than 48% of the travel time inside ORs extracted according to the proposed method. Moreover, results also show that (i) attending to users mobility, ten seconds is the minimum offloading time window that can be considered; (ii) offloading predictive methods can have variable performance according to the period of the day; and (iii) per-user opportunistic decision models can determine offloading system design and performance. This work was published at ACM CHANTS 2018 (jointly with ACM MobiCom)  47. Next we extended the above work as following.

We then assess the mobility predictability in an offloading scenario using theoretical and algorithmic evaluation of several mobility predictors. The results show that mobility predictability for offloading purposes is far more challenging than mobility between PoIs. Here, machine learning (ML) predictors outperform common Markov Chain (MC) predictors used in the literature by at least 15%, revealing the importance of context information in an offloading scenario. The conclusions and findings on offloading mobility properties are likely to generalise for varied urban scenarios given the high degree of similarity between the results obtained for the two different and independently collected mobility datasets.

This last extended work is published at the IEEE Transactions on Network and Service Management  46. The work and the collaboraiton with the PhD Emmanuel Lima is still on-going, a collaboration started when he spent 4 months in 2018 as an intern in our previous team, and his advisors.

The comprehension of preferences related to mobility decisions of an urban population opens new perspectives to tackle the consequences of urbanization. Detecting transportation modes' usability in spatiotemporal urban trajectories enriches such mobility comprehension. With this goal, we introduce Popayi, a transportation mode detection strategy that bases its design on the Ordinal Pattern (OP) transformation applied to mobility-related time series.

Popayi can quantify time-series dynamics in linear time, muscling time series' characteristics that straightforward classification strategies can use. This new strength comes with a low-complex cost, avoiding the need for high computational and methodological complexities in the current Machine Learning (ML) and Deep Learning (DL) literature. Popayi uses polar geodesic representation and amplitude information in time series, bringing the multivariate capability to the standard uni-dimensional OP transformation.

Our experiments show that POPAyI: (i) perfectly adapts to multi-dimensional mobility time series and individuals' natural non-linear mobility behavior. (ii) presents consistent detection results in any considered number of transportation mode's classes with efficiency in terms of storage and computation complexity, using fewer features than ML approaches and computational resources than DL methods. Indeed, Popayi presents classification results equivalent to DL approaches, requiring 10 to 1000 times fewer parameters. For instance, we can increase the F1-score by 2% using 1000 fewer parameters than a lightweight DL approach.

This work has been just accepted to the IEEE Internet of Things Journal (notification received on 6th January 2024).

Due to their complexity and opaqueness, cellular networks have been subject to numerous attacks over the past few decades. These attacks are a real problem to telecom operators and cost them about USD 28.3 Billion annually, as reported by the Communications Fraud Control Association. SIMBox fraud, which is one of the most prevalent of these telephone frauds, is the main focus of this work. SIMBox fraud consists of diverting international calls on the VoIP network and terminating them as local calls using an off-the-shelf device, referred to as SIMBox.

In this work, we first survey both the existing literature and the major SIMBox manufacturers to provide comprehensive and analytical knowledge on SIMBox fraud, fraud strategies, fraud evolution, and fraud detection methods. We provide the necessary background on the telephone ecosystem while extensively exploring the SIMBox architecture required to understand fraud strategies. We provide a complete introductory guide for research on SIMBox fraud and stimulate interest for SIMBox fraud detection, which remains little investigated. We also present insights into tomorrow's SIMBox fraud detection challenges. This survey is published in the IEEE Communication and Tutorial Surveys journal  9 and a technical report can be found in  42.

SIMBox fraud involves diverting international cellular voice traffic from regulated routes and rerouting it as local calls in the destination country. It has significantly affected cellular networks worldwide, generating $\$3.11$ Billion of losses annually and threats to national security. Yet, SIMBox fraud is still an open issue being little addressed in the literature and hardly detected by operators due to two main challenges: (c1) the scarcity of ground-truth-enriched datasets and (c2) the difficulty of leveraging detection. In this work, we introduce the FraudZen framework to tackle (c1) by providing mobile communication datasets (i.e., Charging Data Records/CDRs) with real-world fraudulent ground truth. Furthermore, such CDRs are associated with explicit knowledge of the fraudsters' behavior, i.e., a fraud model, thus filling the gap for tackling (c2). For this, we first identify real-world fraud capabilities via an extensive review of current in-market simbox appliances. We then introd,uce simbox fraud modeling, grasping fraudsters' intents and enabling the design and forecast of such frauds. Such modeling is embedded in the design of FraudZen opne-source simulator, an environment for the scalable simulation of SIMBbox frauds. It is based on the well-known and broadly used LTE-SIM tool in which we added all the required components to simulate SIMBox fraud. Besides, we inserted various traffic generators and realistic mobility modeling, providing lifelike CDR data and ground-truth for comprehensive fraud detection analysis. We validate FraudZen's ability to simulate efficient fraud models and release related generated CDRs datasets. At last, we leverage FraudZen at the in-depth evaluation of literature ML-based fraud detection while considering several fraud- and detection-related parameters. The obtained insights provide detailed hints to future fraud mitigation design.

The FraudZen simulator (cf. Section 6) is mentioned in the Software section and can be found at Inria GitLab. Related publications were published at IEEE WCNC 2023 20, at the ACM Conext Student Workshop 2022  44, at the NetMob 2023 conference, (Book of Abstract) , and at the French Cores 2022  41 and 2023 19. This latter was awarded as the best paper at Cores. An extended version is also under submission.

Cellular SIMBox fraud bypasses international mobile calls and routes them through the internet as local mobile calls in the destination country, using VoIP GSM gateways equipped with multiple SIM cards, also known as "SIMBox." This fraud causes annual financial losses of up to $3.11 billion, national security threats, and phone conversation privacy breaches. Current approaches to mitigate SIMBoxx fraud present open issues that affect their effectiveness. They lack robustness against the constant refinement of fraudsters' strategies or involve a certain implementation complexity that hinders their widespread deployment in operator networks.

This paper presents Sign, a new mitigation approach based on cellular signaling data analysis. Sign is the first-of-the-literature real-time prevention methodology that is beyond fraudster-reach and largely deployable. Sign focuses on the cellular signaling of user devices during the network attachment, aiming to block fraudulent SIMBbox devices before they can connect to the network. Through extensive indoor and outdoor experimentation, we empirically show that fraudulent SIMBox devices cause significant latency than legitimate devices during the network attachment. Especially in the authentication phase, fraudulent SIMBox devices' minimum latency is 23$\times$ higher than their legitimate counterparts. We analyze such latency overhead, showing it is fundamentally shaped by factors beyond fraudsters' control, i.e., LTE standards for authentication and Internet-based communication related protocols and vagaries. Therefore, we propose a SIMBox fraud prevention approach that adapts the standardized authentication procedure at the cellular edge, at no cost for mobile operators.

This work is under submission to a journal.

Cellular networks provide digital communications for more than five billion people around the globe. Besides, their openness to the general public, opaqueness, and complexity have exposed cellular networks to attacks that have tremendously grown over the previous decades. According to the Communication Fraud Control Association's 2021 report, worldwide mobile network operators are experiencing as much as $39.89 billion annually due to illegal activities on their surfaces. Among such illegitimate activities, SIMBox international bypass fraud is one of the most prevalent, having a severe impact manifold.

SIMBox fraud involves diverting international cellular voice traffic from regulated routes and rerouting it as local calls in the destination country from a VoIP-GSM gateway (i.e., SIMBox). Affecting countries worldwide, this problem impairs operators' revenues, network quality, networking research, and national security. Mainly in developing countries, up to 70% of incoming international call traffic is terminated fraudulently. Even worse, SIMBox fraud allows international terrorists to conduct covert activities, masquerading as national subscribers.

In this context, many challenges are added. First, while mobile network datasets (i.e., Charging Data Records or CDRs) are the primary data type leveraged for operators' fraud detection, they are intrinsically private. CDRs hold sensitive information about subscribers' habits, hardening their shareability to the research community and, at the same time, curbing fraud detection investigations. Second, fraudsters' behavior changes over time to adapt to the target solutions, making detection lag behind. In particular, SIMBox fraud increasingly mimics human communication behavior regarding traffic, mobility, and social habits perceptible in CDRs. Third, considering the low related investment, the fraud is quickly profitable. Therefore, the detection time is crucial for effective long-term mitigation.

This thesis tackles international bypass fraud understanding and mitigation while addressing the aforementioned challenges.

Through in-depth evaluations, we validate this thesis's contributions to accomplish a pipeline to handle the fraud: from Fully understanding SIMBox frauds and detection limitations to Long-term fraud mitigation by anticipation and rapid retort.

This topic was addressed at the Anne Josiane's PhD thesis 32 under the supervision of Aline C. Viana and Alain Tchana and under INRIA funding. Anne Josiane defended her Ph.D. on 11th May 2023 and is currently a Post-Doc fellow at TU-BErlin.