IMDEA Networks Institute Digital Repository
Not a member yet
1915 research outputs found
Sort by
On Green Edge Computing with Machine Learning Applications
Get PDFThis thesis finds efficient solutions to dynamically support verticals at the edge while at the same time keeping an eye on sustainability (both in economical and energy-efficiency terms) for network operators. More in detail, this thesis supports novel verticals by developing intelligent algorithms (e.g., heuristics and DRL-based solutions) for the allocation and migration of computing demanding tasks in edge servers dependable on intermittent renewable sources, with the ultimate goal of maximizing revenues for network operators while leveraging as much as possible the presence of renewable sources (i.e., decreasing overall costs).
We first give an updated review of the state-of-the-art, focusing on several important aspects of the Multi-access Edge Computing (MEC) provisioning (such as standardization efforts, techniques to efficiently deploy and support migration of end-user applications, and the offloading of computing tasks). This also gives the motivations behind this thesis’ goals and contributions. Next, we focus on proposing a novel scenario, green edge gaming, where edge computing resources are partially or completely dependent on renewable sources and they have to accommodate heavy computing tasks coming from gaming devices. Another novelty of this scenario is that, since edge servers are located closely, it is possible to migrate allocated gaming jobs between edge servers, according for instance to the availability of green energy.
Next, we leverage powerful machine learning techniques such as Deep Reinforcement Learning (DRL) to propose a DRL-based solution for the allocation and migration of Augmented Reality (AR) tasks at the edge.
Since the goal of maximizing the admittance of AR tasks while leveraging as much as possible the green energy availability is conflicting, we use a proportional fairness structure, which, thanks to the DRL approach, helps to find a sweet spot between these two goals compared to greedy heuristics.
In conclusion, in this thesis, we propose two novel solutions to tackle the problem of allocation and migration jobs in an edge infrastructure, where edge servers depend on intermittent renewable sources. Since one of the key pillars of 6G networks is sustainability, this thesis could lay the foundation for more studies in this evolving scenario.TelematicsUniversidad Carlos III de Madrid, Spai
A prediction-model-assisted reinforcement learning algorithm for handover decision-making in hybrid LiFi and WiFi networks
Get PDFThe handover process in hybrid light fidelity (LiFi) and wireless fidelity (WiFi) networks (HLWNets) is very challenging due to the short area covered by LiFi access points and the coverage overlap between LiFi and WiFi networks, which introduce frequent horizontal and vertical handovers, respectively. Different handover schemes have been proposed to reduce the handover rate in HLWNets, among which handover skipping (HS) techniques stand out. However, existing solutions are still inefficient or require knowledge that is not available in practice, such as the exact user's trajectory or the network topology. In this work, a novel machine learning-based handover scheme is proposed to overcome the limitations of previous HS works. Specifically, we have designed a classification model to predict the type of user's trajectory and assist a reinforcement learning (RL) algorithm to make handover decisions that are automatically adapted to new network conditions. The proposed scheme is called RL-HO, and we compare its performance against the standard handover scheme of long-term evolution (STD-LTE) and the so-called smart handover (Smart HO) algorithm. We show that our proposed RL-HO scheme improves the network throughput by 146% and 59% compared to STD-LTE and Smart HO, respectively. We make our simulator code publicly available to the research community.European Union’s Horizon 2020 Marie Sklodowska Curie grant ENLIGHT’EM (814215)Project RISC-6G, reference TSI-063000-2021-59, granted by the Ministry of Economic Affairs and Digital Transformation and the European Union- NextGenerationEU through the UNICO-5G R&D Program of the Spanish Recovery, Transformation and Resilience Plan.MSCA Postdoctoral Fellowship grant RISA-VLC (101061853)TRUEpu
FedQV: Leveraging Quadratic Voting in Federated Learning
Get PDFFederated Learning (FL) permits different parties to collaboratively train a global model without disclosing their respective local labels. A crucial step of FL, that of aggregating local models to produce the global one, shares many similarities with public decision-making, and elections in particular. In that context, a major weakness of FL, namely its vulnerability to poisoning attacks, can be interpreted as a consequence of the \emph{one person one vote} (henceforth \emph{1p1v}) principle that underpins most contemporary aggregation rules.
In this paper, we introduce \textsc{FedQV}, a novel aggregation algorithm built upon the \emph{quadratic voting} scheme, recently proposed as a better alternative to \emph{1p1v}-based elections. Our theoretical analysis establishes that \textsc{FedQV} is a truthful mechanism in which bidding according to one's true valuation is a dominant strategy that achieves a convergence rate matching that of state-of-the-art methods. Furthermore, our empirical analysis using multiple real-world datasets validates the superior performance of \textsc{FedQV} against poisoning attacks. It also shows that combining \textsc{FedQV} with unequal voting ``budgets'' according to a reputation score increases its performance benefits even further. Finally, we show that \textsc{FedQV} can be easily combined with Byzantine-robust privacy-preserving mechanisms to enhance its robustness against both poisoning and privacy attacks.the Ministry of Economic Affairs and Digital Transformation and the European Union NextGenerationEU/PRTRTRUEinpres
Metric meta-learning in deep learning models for intent-based networking
Get PDFThe evolution towards self-configuring communication networks necessitates anticipatory approaches for network management, aiming for zero-touch orchestration. Intent-Based Networking (IBN) represents a transformative perspective by automating the translation of user-defined intents into effective network operations. Rather than specifying how tasks should be executed, users only have to define desired outcomes using high level human intents, such as secure data transfer between two locations, and the IBN system will adapt itself accordingly. The system continuously monitors and adjusts settings to ensure these goals are met, simplifying network management, enhancing agility, and speeding up service deployment. By abstracting complex configurations, IBN offers a more intuitive, efficient, and business-aligned way to manage networks, leveraging automation, artificial intelligence, and machine learning to align network configurations with user-specified outcomes.
To achieve this, network orchestrators must use automated decision models that accurately actualize intent-based objectives, especially in anticipatory networking where predicting the relationship between management decisions and performance is often impossible. Traditional prediction models, restricted by generic and manually set objectives, require perfect knowledge of this relationship, which is not feasible for many tasks. For instance, while it is possible to measure the resulting performance of a resource allocation decision a posteriori, determining the outcome in advance is extremely difficult. This highlights the need for advanced predictive models that autonomously understand and adapt to the interplay between predictions and targeted network management objectives. This thesis introduces a model designed to bridge the gap between loss functions and performance metrics in regression tasks, representing a significant advancement in automated machine learning. Real-world examples underscore the model's effectiveness, aligning predictions with the detailed objectives of IBN and showcasing its adaptability and capability to meet the demands of anticipatory networking tasks.
Central to this, is the goal of addressing the \say{loss-metric mismatch} problem in machine learning, where the loss function used for training deep neural networks (DNNs) does not align with the actual performance metric or the translated anticipatory Management and Network Orchestration (MANO) objective in the context of IBN. The first part of the thesis delves into this issue, proposing a model that aligns training objectives with performance metrics, enhancing the practical application of DNNs in network management. However, in the rapidly evolving landscape of network management, another critical aspect is aligning disparate entities towards a common objective. Traditional methodologies, which prioritize individual optimization goals, often lead to competitive behaviors that can deteriorate overall system performance. This is particularly problematic in complex scenarios requiring a collaborative and holistic approach. To address this, the second part of the thesis builds on the initial model, introducing an improved version that coordinates multiple concurrent predictions to achieve a common goal with minimal assumptions. This model maintains a solution to the loss-metric mismatch while extending its application to multiple predictors, significantly widening the range of potential network management applications. Exhaustive testing in controlled environments and real-world applications demonstrates the superior performance of this approach, meeting the intricate demands of anticipatory IBN.
The final part of the thesis is dedicated to exploring the explainability of DNN models. In today's world, where ethical considerations and enhanced understanding are paramount, explainability is crucial for comprehending a model's behavior and discerning normal functioning from potential malfunctions, whether due to inherent flaws or malicious data attacks. The ability to explain a model's decisions and processes is not just a technical requirement but also an ethical imperative, ensuring transparency and trust in automated systems. Special attention is given to the significance of explainability in spatio-temporal models used for detecting potential attacks. The thesis examines the explainability aspects of the models proposed above, demonstrating their inherent transparency compared to other techniques like Reinforcement Learning (RL) thanks to their inherent architectures. This exploration is vital for providing insights into the model's decision-making processes, enhancing trustworthiness, and facilitating the identification and mitigation of vulnerabilities. By focusing on explainability and interpretability, the thesis aims to contribute to the development of more transparent, reliable, and ethically sound machine learning models, particularly in network management and security.
Overall, this thesis presents a comprehensive exploration of advanced network management strategies. It identifies and addresses critical challenges in aligning the predictive capabilities of machine learning models, particularly DNNs, with the nuanced objectives of IBN. The proposed models, which resolve the \say{loss-metric mismatch} and ensure unified goal achievement, mark significant advancements in automated network management. These models are not only theoretically relevant but also validated through extensive real-world applications, demonstrating their practical efficacy and adaptability in dynamic networking environments. Furthermore, the thesis emphasizes the importance of explainability in machine learning, highlighting its crucial role in ensuring the reliability, transparency, and ethical integrity of automated systems. By delving into the explainability of DNN models, the research significantly contributes to the development of more transparent and trustworthy machine-learning solutions, offering a robust framework for the future of network management.TelematicsIMDEA Network
An In-Depth Analysis of Advanced Time Series Forecasting Models for the Open RAN
Get PDFForecasting is instrumental to efficiently manage network resources. In this workshop paper, we make the following contributions. First, we carry out the first assessment of recently proposed advanced forecasting techniques by the AI community, namely DLinear and PatchTST, when applied to the prediction of mobile traffic load and number of users connected to a single Base Station (BS ). We compare these techniques with the well-known Long-Short Term Memory (LSTM) models that are widely adopted in mobile network tasks. Second, we analyze the accuracy tradeoff of these Artificial Intelligence (AI) techniques for single-
and multi-step prediction horizons. Third, we profile the operation of all these black-box predictors with an EXplainable Artificial Intelligence ( XAI) lens by using AIChronoLens, a new tool that links legacy XAI explanations with the temporal properties of the input sequences. We find that DLinear excels in single-step horizon predictions while PatchTST and LSTM are more accurate in multi-step horizon predictions. Our XAI study reveals that, unlike PatchTST and LSTM, DLinear focuses its prediction decisions
on a few key samples of the input sequences, which ultimately lets it match the ground truth closely.TRUEinpres
Reviewing War: Unconventional User Reviews as a Side Channel to Circumvent Information Controls
No full textDuring the first days of the 2022 Russian invasion of Ukraine, Russia's media regulator blocked access to many global social media platforms and news sites, including Twitter, Facebook, and the BBC. To bypass the information controls set by Russian authorities, pro-Ukrainian groups explored unconventional ways to reach out to the Russian population, such as posting war-related content in the user reviews of Russian businesses available on Google Maps or Tripadvisor. This paper provides a first analysis of this new phenomenon by analyzing the unconventional strategies used to avoid state censorship in the Russian Federation during the conflict. Specifically, we analyze reviews posted on these platforms from the beginning of the war to September 2022. We measure the channeling of war-related messages through user reviews on Tripadvisor and Google Maps. Our analysis of the content posted on these services reveals that users leveraged these platforms to seek and exchange humanitarian and travel advice, but also to disseminate disinformation and polarized messages. Finally, we analyze the response of platforms in terms of content moderation and their impact.TRUEpu
Demo: Streaming Video over 360 Degrees Visible Light Communication
Get PDFIn this paper we present the demo of a 360 degrees visible light communication (VLC) transmitter built over the low-cost OpenVLC platform. We present how we can stream a video broadcasted from a VLC transmitter providing large angular coverage, and displaying it seamlessly in the receiver. We show how the setup is implemented and evaluate its performance. This work shows the feasibility of low-cost 360 degrees VLC transmitters complying with real-worldrequirements of distance, coverage angle, mobility, data rate and robustness against background light.Technology Innovation InstituteTRUEpu
Differences in the Toxic Language of Cross-Platform Communities
Get PDFCross-platform communities are social media communities that have a presence on multiple online platforms. One active community on both Reddit and Discord is dankmemes. Our study aims to examine differences in harmful language usage across different platforms in a community. We scrape 15 communities that are active on both Reddit and Discord. We then identify and compare differences in type and level of toxicity, in the topics of the harmful discourse, in the temporal evolution of toxicity and its attribution to users, and in the moderation strategies communities across platforms.
Our results show that most communities exhibit differences in toxicity depending on the platform.
We see that toxicity is rooted in the different subcultures as well as in the way in which the platforms operate and their administrators moderate content. However, we note that in general terms Discord is significantly more toxic than Reddit. We offer a detailed analysis of the topics and types of communities in which this happens and why, which will help moderators and policymakers shape their strategies to mitigate the harm on the Web. In particular, we propose practical and effective strategies that Discord can implement to improve their platform moderation.UK's Research centre on Privacy, Harm Reduction & Adversarial Influence onlineSpanish Ministry of Science and InnovationESF Investing in your futureEuropean Union-NextGenerationEUTRUEinpres
ReviewingWar: Unconventional User Reviews as a Side Channel to Circumvent Information Controls
No full textDuring the first days of the 2022 Russian invasion of Ukraine,
Russia’s media regulator blocked access to many global social
media platforms and news sites, including Twitter, Facebook,
and the BBC. To bypass the information controls set by
Russian authorities, pro-Ukrainian groups explored unconventional
ways to reach out to the Russian population, such
as posting war-related content in the user reviews of Russian
businesses available on Google Maps or Tripadvisor. This paper
provides a first analysis of this new phenomenon by analyzing
the unconventional strategies used to avoid state censorship
in the Russian Federation during the conflict. Specifically,
we analyze reviews posted on these platforms from
the beginning of the war to September 2022. We measure the
channeling of war-related messages through user reviews on
Tripadvisor and Google Maps. Our analysis of the content
posted on these services reveals that users leveraged these
platforms to seek and exchange humanitarian and travel advice,
but also to disseminate disinformation and polarized
messages. Finally, we analyze the response of platforms in
terms of content moderation and their impact.Horizon 2020Agencia Estatal de InvestigaciónTRUEinpres
HyperGraphDis: Leveraging Hypergraphs for Contextual and Social-Based Disinformation Detection
Get PDFIn light of the growing impact of disinformation on social, economic, and political landscapes, accurate and efficient identification methods are increasingly critical. This paper introduces HyperGraphDis, a novel approach for detecting disinformation on Twitter that employs a hypergraph-based representation to capture (i) the intricate social structures arising from retweet cascades, (ii) relational features among users, and (iii) semantic and topical nuances. Evaluated on four Twitter datasets -- focusing on the 2016 U.S. Presidential election and the COVID-19 pandemic -- HyperGraphDis outperforms existing methods in both accuracy and computational efficiency, underscoring its effectiveness and scalability for tackling the challenges posed by disinformation dissemination. HyperGraphDis displays exceptional performance on a COVID-19-related dataset, achieving an impressive F1 score (weighted) of approximately 89.5%. This result represents a notable improvement of around 6% compared to existing methods. Additionally, significant enhancements in computation time are observed for both model training and inference. In terms of model training, completion times are accelerated by a factor ranging from 2.3 to 9.3 compared to previous methods. Similarly, during inference, computation times are 1.3 to 7.2 times faster than the state-of-the-art.IMDEA NetworksTRUEinpres