LLM inference exhibits substantial variability across queries and execution phases, yet inference configurations are often applied uniformly. We present a measurement-driven characterization of workload heterogeneity and energy-performance behavior of LLM inference under GPU dynamic voltage and frequency scaling (DVFS). We evaluate five decoder-only LLMs (1B-32B parameters) across four NLP benchmarks using a controlled offline setup. We show that lightweight semantic features predict inference difficulty better than input length, with 44.5% of queries achieving comparable quality across model sizes. At the hardware level, the decode phase dominates inference time (77-91%) and is largely insensitive to GPU frequency. Consequently, reducing GPU frequency from 2842 MHz to 180 MHz achieves an average of 42% energy savings with only a 1-6% latency increase. We further provide a use case with an upper-bound analysis of the potential benefits of combining workload-aware model selection with phase-aware DVFS, motivating future energyefficient LLM inference systems.
Federated Learning (FL) is an emerging paradigm for decentralized training of Machine Learning models. It has been the subject of a large corpus of research due to its innovative approach to handling sensitive data. A common practice in the FL literature is to run simulations on a single compute node to assess the performance of FL algorithms. While simulation enables fast prototyping and validation of algorithmic concepts, it may face limitations in reproducing the real system's performance in heterogeneous environments such as the Computing Continuum, and particularly on resource-constrained Edge devices. Conversely, emulation on distributed testbeds offers more effective means to accurately reproduce the performance of real-world devices. However, to the best of our knowledge, no prior research has investigated the differences between simulation and emulation in FL experiments. In this paper, we study the complementarity of these approaches and discuss their respective challenges, as a first step towards reproducibility of FL experiments. We illustrate our study with a real-life application used as a baseline: an outdoor air quality forecasting framework with real-world sensors. Our results show that simulation can be used to accurately reproduce model performance metrics, while emulation can effectively reproduce the system performance of real-world experiments. Finally, we present a set of lessons learned on the challenges of FL reproducibility and the selection of experimental infrastructures for FL experiments and applications.
Federated learning (FL) is a distributed machine learning paradigm enabling collaborative model training while preserving data privacy. In today's landscape, where most data is proprietary, confidential, and distributed, FL has become a promising approach to leverage such data effectively, particularly in sensitive domains such as medicine and the electric grid. Heterogeneity and security are the key challenges in FL, however, most existing FL frameworks either fail to address these challenges adequately or lack the flexibility to incorporate new solutions. To this end, we present the recent advances in developing Appfl, an extensible framework and benchmarking suite for federated learning, which offers comprehensive solutions for heterogeneity and security concerns, as well as user-friendly interfaces for integrating new algorithms or adapting to new applications. We demonstrate the capabilities of Appfl through extensive experiments evaluating various aspects of FL, including communication efficiency, privacy preservation, computational performance, and resource utilization. We further highlight the extensibility of Appfl through case studies in vertical, hierarchical, and decentralized FL. Appfl is fully open-sourced on Github at https://github.com/APPFL/APPFL.
We are witnessing a tremendous expansion in computational power, leading to monumental technological advances with various practical applications. The need for parallelism originated from the limitations of single-processing performance. Increasing the transistor density on integrated circuits was the way to go, thus creating faster and more complex monolithic processors. However, these circuits approached their limits of hardware expansion, so the single-processor model has become unreliable and insufficient overall. This is how multicore architectures were born and, with them, the notion of parallel computing. Our research background in the Drop Computing paradigm and Interest in studying parallel programming paradigms encouraged us to propose a new approach towards parallel processing in mobile, ad-hoc, opportunistic networks. This paper adapts one fundamental message-passing standard for parallel architectures within the Drop Computing context. We define a new library DroMPI. Besides the challenges of parallel programming, the solution has to address the challenges imposed by hardware constraints, limited resources, and the decentralized model specific to Drop Computing.
Current security cloud practices can successfully protect stored data and data in transit, but they do not keep the same protection during data processing. The data value extraction requires decryption, creating critical exposure points. As a result, privacy-preserving techniques are emerging as a crucial consideration in cloud computing. The homomorphic processing of machine learning models in the cloud represents a central challenge. The activation function is fundamental in constructing a privacy-preserving Neural Network (NN) with Homomorphic Encryption (HE). Standard activation functions require operations not supported by HE, so it is necessary to find cryptographically compatible replacement functions to operate over encrypted data. Multiple approaches address the limitation of function compatibility with polynomial approximation. These functions should exhibit a trade-off between complexity and accuracy, limiting the efficiency of conventional approximation techniques. The current literature on polynomial approximation of NN activation functions still lacks a thorough review. In this paper, we comprehensively review the standard activation functions of modern NN models and current polynomial approximation approaches. We highlight fundamental features to consider in the activation function and the approximation technique to operate over encrypted data.
For microservices-based real-time stream processing applications, computing at the edge delivers fast responses for low workloads, but as workload increases, the response time starts to slow down due to limited compute capacity. Abundant compute capacity in the cloud delivers fast responses even for higher workloads, but incurs very high cost of operation. For applications which can tolerate latencies up to a certain limit, using either of them has one or the other drawback and for different applications and edge infrastructures, it is non-trivial to decide when to use only edge resources and when to leverage cloud resources. In this paper, we propose CLAP, which dynamically understands the relationship between workload and application latency, and automatically adjusts placement of microservices across edge and cloud computing continuum, with the goal of jointly reducing latency as well as cost of running microservices-based streaming applications. CLAP leverages Reinforcement Learning (RL) technique to learn the optimal placement for a given workload and based on the learnings, adjusts placement of microservices as the application workload changes. We conduct experiments with real-world video analytics applications and show that CLAP adapts placement of microservices in response to varying workloads and achieves low latency for applications in a cost-efficient manner. Particularly, we show that for two real-world video analytics applications i.e. human attributes and face recognition, CLAP is able to reduce average cost (across 4 days at different locations) by 47% and 58% for human attributes detection and face recognition application, respectively, while consistently maintaining latency below the tolerable limit.
Cybersecurity efforts aim to protect computer systems, especially complex systems with dynamically changing architectures like mobile computational clouds, from digital attacks. The use of machine learning (ML) models in these systems to analyze and process data, allocate resources or monitor and manage infrastructure is increasingly the field of hackers and the target of sophisticated and increasingly difficult-todetect attacks. Adversarial machine learning (AML) is a class of techniques and methods that attempt to "fool" classical ML algorithms by using, among others, deceptive data or modifying the parameters of cyber-physical systems (CPS).This paper presents a new generic evasion attack model for generating targeted attacks on the ML binary classifiers in domains with nonlinear constraints. We used GD (Gradient Descent) and Adam (Adaptive Moment Estimation) optimization algorithms as the most important internal components of two variants of our attack model. Such a model was then evaluated in an experimental analysis. The achieved results show that the new generic evasion attack model effectively generates targeted attacks on ML binary classifiers, utilizing GD and Adam optimization algorithms. This model was evaluated through experimental analysis, demonstrating its potential in cybersecurity contexts. Through this approach, the research contributes significantly to the field of adversarial machine learning by presenting a novel method that challenges the robustness of ML models under specific nonlinear constraints.
Polar science is an umbrella term for several research disciplines practiced in Earth’s polar regions and other planets. Polar science research includes studies performed on land (such as geology and archeology on the circumpolar tundra), air (atmospheric research), sky (planetary observations due to dark, clear, and clean skies during the polar nights), and ocean (marine studies in the Arctic and the Southern Oceans). Many polar science research works look into overlapping complementary studies. Distributed computing is a discipline of computer science that involves computations across multiple distinct physically separated computing resources, such as computer clusters, clouds, and the edge. Distributed computing has enabled efficient, scalable, and elastic execution of complex computational problems on utility hardware without expecting access to expensive infrastructure or supercomputers.In this paper, we screen 770 for the interdisciplinary research of distributed computing used in or developed for polar science. After systematically removing the irrelevant studies, we assess the full text of 72 papers to understand the distributed systems research landscape for polar science. We then specifically study 22 polar science research works that develop or heavily utilize distributed computing frameworks or principles in detail. Our study finds distributed execution frameworks instrumental for polar science due to the complex and real-time computational needs, coupled with the remote location requiring efficient network bandwidth availability and usage to access remote resources.
This paper explores the application of machine learning for practical applications in the context of Beyond 5G (B5G) communications. A variety of machine learning techniques, including neural networks, was applied on a labeled dataset about network slicing. Neural network models demonstrate superior performance in optimizing virtual network slices, crucial for enhancing Internet of Things (IoT) connectivity and efficiency. The findings can assist telecommunications professionals and policymakers, offering practical perspectives on AI technologies that can be applied in B5G scenarios for large communications networks.
The increasing energy consumption in cloud-based systems has led to concerns about the environmental impact and cost-effectiveness of such cloud-based systems. To address this issue, researchers have proposed various techniques for improving the energy efficiency of cloud-based systems, including load balancing. Geographical load balancing is a unique technique that distributes the workload across data centers based on their proximity to users, thus reducing the distance that data must travel and minimizing energy consumption. This paper presents a unique geographical load balancing system Durga, that is proven to be efficient and energy-saving. First, the paper comprehensively reviews the literature on geographical load balancing, including its implementation, benefits, and limitations. Then, we design an algorithm that reduces the time to resolve the target data center location to serve the request. In this way, it becomes easier for the data center engineers to route the data packets efficiently reducing the time to access data and thus saving energy. To make this research as realistic as possible, we used Apache Jmeter to simulate real-life scenarios. Finally, using the Haversine formula, we explore Orthodromic Distance and calculate the distance between the user and the data center location. The findings suggest that geographical load balancing can significantly reduce the energy consumption of cloud-based systems and improve their overall performance. The paper also highlights challenges and future research directions, including the need for more advanced algorithms to optimize load balancing and minimize energy consumption.
With the advancement of technology and the spread of multi-core systems, the need for parallelization arises and the interest in programming models is growing. At the same time, new distributed computing models have been proposed, being in fierce competition to obtain the highest possible performance. The Drop Computing Paradigm proposes the idea of decentralized computing over ad-hoc opportunistic networks of mobile and Edge devices. In this respect, the Drop Computing model does not only aim to achieve a minimum turnaround time but also to optimize other characteristics related to mobile devices, such as limited resources and opportunistic communication. Therefore, it is necessary to define a new programming model called DroMPI that intends to extend the capabilities of current parallel and distributed programming models, based on the Drop Computing paradigm. Therefore, the solution aims to develop a library that takes advantage of hardware capabilities in the interest of the Drop Computing paradigm and also provides programmers with a high-level programming interface. The library’s features will be based on the Message Passing Interface (MPI) standard, which will be responsible for inter-node parallelization. The name of the library, DroMPI, is an acronym for Drop Computing and MPI. The implementation of the model will be responsible for the management of communication between nodes and for providing an Application Programming Interface (API) for the development of parallel applications in the Drop Computing paradigm.
As Internet of Things (IoT) devices become more attractive attack targets, cyber threat intelligence would require accurate information about these smart gadgets within a city or a region of interest. Yet, collecting the security status (e.g., whether the latest security patches have been installed) of the deployed IoT devices at a large scale remains challenging, due to the lack of technical capability. In this work, we propose and develop a technical solution to this problem-we propose IoTell for enabling local regulators to monitor the security status of deployed IoT devices. With IoTell in place, a security regulator receives accurate, periodically-updated (e.g., every day) counts of all IoT firmware versions operating in a region (e.g., a city). A naive approach to IoTell would easily violate user privacy and become vulnerable to data manipulation. In this work, we present an end-to-end privacy-preserving protocol architecture for IoTell that addresses the potential privacy and security risks. IoTell requires only widely available market-ready technical capabilities and minimal addition to IoT devices and network operators.
Understanding user behavior and anticipating user decisions is difficult. Human Machine Interfaces (HMIs) provide guiding principles for delivering a seamless interface between users, machines, and infrastructure for Intelligent Transportation Systems (ITS). User behavior comprehension and prediction is unique for every user and involves understanding several dynamic factors. Research has been done leveraging machine learning (ML) and artificial intelligence (AI) techniques to address this interesting interplay by training Neural Network (NN) models to predict the user’s next action based on various input features. The proposed method is to employ Reduced Canonical Form (RCF), a mathematical framework, to represent and simplify user behavior parameters, which enables the production of more intuitive and responsive HMIs for ITS. After representing the user behavior model using RCF, to prove our hypothesis, we have introduced mathematical proofs to show the effectiveness and validity of RCF in order to map user behavior efficiently.
In this paper, we discuss Zerologon, a critical vulnerability rated at 10 in severity by authorities, and identified as one of the most exploited vulnerabilities in recent years, as per research. It enables a malicious actor to impersonate any computer or root domain controller, thereby infiltrating the network. In other words, an attacker can gain authentication without needing credentials. This paper not only provides background information on the vulnerable system, including a description of the domain controller and remote procedure, but also offers an in-depth analysis of the potential attack methods employed by malicious actors and their resulting impacts. To facilitate better comprehension, we include relevant mathematical formulas and cryptography knowledge in this section. The novelty of this paper primarily lies in the detailed explanation of the three phases of the solution. The first phase is the initial deployment phase; the second is the finding phase; the third is the ’Enforcement Phase.’ This study aims to investigate and analyze the Zerologon vulnerability, focusing on how such attacks infect networks, impact companies, and can be deterred or mitigated.
This paper proposes a novel framework for an Internet of Things (IoT) testbed dedicated to recording data on smart home activity and facilitating the development and evaluation of AI-powered security solutions for smart infrastructure. The framework addresses the critical need for accessible and standardized threat detection data in smart spaces, promoting progress in the security of smart homes. This paper presents a framework for the creation of a standardized smart home threat detection data set repository for a controlled testbed that simulates smart home environments, equipped with IoT sensors. The proposed framework will serve as a resource for researchers and industry alike, both of whom need sensor-based data to train and test AI models.
The success of the Fifth Generation (5G) cellular network has opened up opportunities for connecting multiple devices and accommodating different services and applications. However, the architecture of 5G and network functions are susceptible to various cyber-attacks and security risks. Threat modeling facilitates the identification of threats, assessment of potential attacks, and the assessment of applying the mitigation targeting a system. While existing research on 5G threat modeling focuses solely on public network attacks targeting the 5G infrastructure, it is equally important to develop tailored approaches for addressing emerging threats occurring at 5G and B5G infrastructure. Furthermore, current research lacks detailed information on threat vectors, entry points, and network vulnerabilities. To address this gap, this paper proposes a threat modeling strategy specifically designed for characterizing and modeling threats in 5G and B5G networks. The proposed XML schema provides information on vulnerabilities, entry points, attacks, TTPs, and correlates countermeasures with the MITRE framework. The proposed approach is exemplified via a DNS amplification attack that employs tactics and techniques relevant to telecommunication networks.
Crowd or mass gatherings at various venues such as entertainment events or transportation systems are faced by individuals on a daily basis. Crowd management for large-scale planned events is essential, and depends on behavioural and mathematical modelling to understand how crowds move in different scenarios. In this paper, we propose the creation of a set of collaborative and services, the CrowdMesh platform, to allow for processing, analysis, and sharing of crowd monitoring datasets constructed over the Edge-Cloud Continuum. Comprehensive tools and best practices are needed to ensure that analysis pipelines are robust, and that results can be replicated.
The rapid advancement of technology has brought about profound changes in various spheres of our society. Among these improvements, the Internet of Things (IoT) stands out as a transformative force, revolutionizing visibility across diverse domains. By furnishing real-time data and insights, IoT empowers businesses to refine their operations and significantly enhance overall efficiency. However, the decentralized nature of IoT architecture presents a substantial challenge in terms of security. As technology proliferates, security emerges as a paramount concern, particularly in safeguarding IoT devices and the data they generate against unauthorized access, data breaches, and potential harm. Simultaneously, the imperative for real-time data streaming underscores the need for a rapid dissemination protocol as a fundamental requirement. Our proposed solution adeptly addresses both of these critical aspects by integrating BDLS, a promising Byzantine Fault Tolerance (BFT) protocol, with Hyperledger Fabric for IoT devices. This integration not only enhances security but also ensures the rapid and secure dissemination of data. Through rigorous evaluation, our study demonstrates the outstanding performance of BDLS Fabric integration in terms of throughput (TPS), surpassing CFT-type protocols and numerous other proposed solutions utilizing PBFT, thereby reinforcing the importance of security in IoT deployments. Moreover, by securing the edge servers and the cloud replica using Hyperledger Fabric-BDLS integration, our solution fortifies the IoT ecosystem against potential threats, establishing a robust foundation for future innovations in distributed ledger technology.
In early 2021, a file containing an organized, interactive dataset of 3 billion email and password combinations was posted on a hacker forum. This dataset, known as the ’Compilation of Many Breaches’ (COMB), represents the largest and most recent data breach compilation to spread on the internet, comprising data from various company breaches. This paper aims to dissect the architecture of COMB, explore its contents, and assess the security dangers it poses. In this study, COMB is analyzed in relation to other data breach compilations, uncovering concerning patterns in the lifecycle of these data files. We find that the risk posed by cybersecurity attacks, such as credential stuffing and deep learning password cracking, escalates significantly when these attacks utilize data breach compilations. Despite the absence of prior academic literature on data breach compilations, the evolution of these compilations and the hazards they create highlight the necessity for further investigation. This study emphasizes the importance of implementing defense solutions, such as proper password hygiene, at both individual and organizational levels to mitigate the damage of COMB and prevent similar future incidents.
The Fintech industry represents the convergence of finance and technology through innovative digital solutions from mobile banking to cryptocurrency. However, it faces big challenges with keeping data safe and systems strong. Traditional methods struggle to keep pace with sophisticated threats and complexities inherent in modern Fintech ecosystems. This paper proposes an approach to address these challenges using Generative AI and blockchain integration to make Fintech systems more resilient. Advanced machine learning algorithms detect and prevent data tampering in the proposed systems. Generative AI is used for threat detection, anomaly recognition, and real-time monitoring in system security. Then, we integrate blockchain technology to enhance the overall resilience of systems. Blockchain technology enhances the reliability of financial services in secure transactions, validating blocks, and distributing control across decentralized networks. These combined methodologies address the critical challenges of data integrity, security, and system resilience in dynamic Fintech systems. The performance analysis demonstrates the efficacy of our proposed framework in enhancing data integrity, security measures, and system resilience within Fintech systems.
