The 1st International Workshop on AI/ML for Edge/Fog Networks (A4E 2022)

Session A4E-OS

Opening Session

10:00 AM — 10:10 AM EDT
May 2 Mon, 10:00 AM — 10:10 AM EDT

Session A4E-KS

Keynote Session

10:10 AM — 10:40 AM EDT
May 2 Mon, 10:10 AM — 10:40 AM EDT


Milind Tambe (Harvard University)

This talk does not have an abstract.

Session A4E-TS1

Technical Session I

10:40 AM — 12:00 PM EDT
May 2 Mon, 10:40 AM — 12:00 PM EDT

MERIT: Multi-Itinerary Tourist Recommendation Engine for Industrial Internet of Things

Abhishek Majumder (Tripura University, India); Joy Lal Sarkar (Research Fellow, Tripura University, India); Bibudhendu Pati (Rama Devi Women's University); Ramasamy V (M.Kumarasamy College of Engineering, Karur, Tamilnadu, India.); Chhabi Rani Panigrahi (Rama Devi Women's University, Bhubaneswar, India); Sudipta Roy (Assam University, India); Vikas Kumar (University of Delhi)

The expansion and usage of the Internet of Things (IoT) in commercial industries and technologies are referred to as the Industrial Internet of Things (IIoT). The continuous collection and transmission of data amongst electronic gadgets provide multiple possibilities for companies and businesses to expand. Tourists flying across the city will learn from the IIoT in a variety of ways. Tourists face a variety of challenges when choosing appropriate tours from a variety of itineraries based on their preferences and restrictions. In this paper, we suggest the MERIT algorithm. It proposes several itineraries which rely on tourists' interests in a given location, weather conditions, the familiarity of itineraries as well as the expense of itineraries. Different sensors linked in the city track the various tourist events, and they are forwarded to a centralized database to be analyzed afterward to an intelligent gadget that will assist to offer better tourist recommendations. We also took into consideration the suggested approach whenever a tourist wishes to explore unfamiliar places, utilizing the Flickr Data Collection to suggest several itineraries. The results indicate that the suggested MERIT approach exceeds the benchmark models based on the real-life matrices.

Secrecy Rate Maximization for THz-Enabled Femto Edge Users using Deep Reinforcement Learning in 6G

Ishan Budhiraja (Bennett University, Greater Noida, India); Himanshu Sharma (Thapar Institute of Engineering and Technology, India); Neeraj Kumar (Thapar University Patiala, India); Raj Tekchandani (Thapar University, India)

Dense deployment of femtocells in heterogeneous networks (HetNets) is critical for satisfying end-user quality-of-service (QoS) requirements. Femtocells can improve the network spectral efficiency and reduce power consumption. However, due to the distributed and dynamic nature, femto edge users (FEUs) cannot resist the attacks of eavesdroppers. Hence, it is essential to enhance the secrecy rate of FEUs for the secure and reliable transmission of data. Aiming to improve the secrecy rate of the FEUs, this paper formulates the power control problem of terahertz (THz)-enabled femto base station (FBS) consisting of multiple FEUs and a single eavesdropper. Since the system is very complex and dynamic, addressing the non-convex optimization problem is difficult, so firstly we use the Markov decision process (MDP) to translate the optimization problem into a multi-agent deep reinforcement learning (DRL) problem. Then, to solve the power control problem, we have used multi-agent Q-learning to enhance the learning ability and reduce the output dimension. Then, we have used deep deterministic policy gradient (DDPG) to convert the policy into a deterministic one and to achieve efficient power control. Simulation results show that the proposed DRL based technique significantly improves the average secrecy rate of FEUs by 16.67% and 5% as compared to existing state-of-art schemes.

Optimal Feature Set Selection for IoT Device Fingerprinting on Edge Infrastructure using Machine Intelligence

Sarvesh Wanode (BITS Pilani, Hyderabad Campus Hyderabad); Milind Anand (BITS Pilani, Hyderabad Campus, Hyderabad, India); Barsha Mitra (BITS Pilani, Hyderabad Campus, India)

The widespread deployment of IoT devices globally has spurred the need for protecting these devices from various cyber security attacks. IoT device fingerprinting is the method of identifying IoT devices by analyzing network traffic. Fingerprinting helps to determine if devices have been subverted by
an attacker by separating anomalous behavior from normal behavior. This kind of device identification is carried out on some edge device with the help of machine learning. With the growing number of IoT devices and the large volumes of traffic generated by them, it is essential to determine an optimal feature
set for device classification. An optimal feature set not only helps to identify the most important features which make maximal contribution in creating a device fingerprint, but also makes the classification model light weight and suitable for deployment on edge devices with low computational power. In this paper, we
present a feature reduction method using three popular machine learning techniques and apply it for classifying IoT devices. Our feature reduction method identifies the most important features by isolating them from the non-essential ones, thereby giving a reduced feature set that can provide a classification accuracy comparable to the original feature vector. Moreover, we use the reduced feature set thus obtained to identify new IoT devices introduced in the network. We have performed experiments using an open source IoT device dataset. The experimental results show that we are able to identify the optimal features that constitute only 19% of the original feature vector. Moreover, the absence of 81% of the initial features does not compromise the performances
of device classification as well as new device identification.

Edge Intelligence-based Privacy Protection Framework for IoT-based Smart Healthcare Systems

Mahmuda Akter (University of New South Wales, Australia); Nour Moustafa (University of New South Wales at Canberra, Australia); Timothy Lynar (University of New South Wales at the Australian Defence Force Academy, Australia)

Federated Learning (FL) mechanisms determine the implications of sensitive data for constructing on-device Machine Learning (ML) to achieve personalisation in a smart application network, for example, sharing critical information of the smart healthcare industry over the Internet of Things (IoT) systems. The main function of centralised FL can be combined with Edge Intelligence (EI) for processing before final aggregation to reduce data manipulation and privacy hazard. However, executing EI in an Edge Computing (EC) layer also poses privacy risks to clients. Differential Privacy (DP) offers a viable solution by adding artificial noise to a parameter before aggregation. This paper introduces a Federated Edge Aggregator (FEA) framework with DP for safeguarding the high-tech healthcare industry using IoT systems. An iteration-based converged Convolutional Neural Network (CNN) model at Edge Layer (EL) is developed to perform EI to balance FL's privacy preservation and model performance over an IoT network. The results demonstrated a 90% accuracy performance after specific iterations, better than those of other baseline approaches with accuracy levels of approximately 80% with the same epsilon value of 4. Also, this framework is faster and more successfully meets the privacy preservation paradigm.

Session A4E-TS2

Technical Session II

12:40 PM — 2:00 PM EDT
May 2 Mon, 12:40 PM — 2:00 PM EDT

A Novel IoT-based Edge Sensing Platform for Structure Health Monitoring

Igor Bisio, Chiara Garibotto, Fabio Lavagetto and Andrea Sciarrone (University of Genoa, Italy)

The emerging framework of 5G, together with the Edge Computing (EC) paradigm, has enabled a number of smart applications in many different scenarios. In the context of smart cities and intelligent transportation systems, Structure Health Monitoring (SHM) has gained increasing attention lately, due to the critical conditions of specific civil infrastructures, which led to unfortunate accidents. In this connection, we propose a low-cost, versatile and self-sustaining platform based on the Narrow Band-IoT (NB-IoT) technology aimed at efficiently and continuously monitoring the conditions of critical civil structures. We also provide both lab and in-site experimental tests comparing the performance obtained with our prototype with respect to professional monitoring equipment. Results show a winning trade-off between performance and unquestionable advantages in deployment, use, maintenance and costs.

Blockchain and Edge Intelligence-based Secure and Trusted V2V Framework Underlying 6G Networks

Sudeep Tanwar (Institute of Technology Nirma University Ahmedabad Gujarat, India); Nilesh Jadav and Rajesh Gupta (Institute of Technology, Nirma University, India)

There is a significant rise in vehicle-to-vehicle (V2V) communication for intelligent transportation systems such as reducing road accidents, traffic congestion, and optimal route planning. The main objective of the V2V communication is to provide real-time monitoring data from vehicles sensors to other vehicles. However, the attackers can exploit this communication by forging the controller area network (CAN) protocol and injecting malicious traffic. In this context, the vehicles mislead by false update messages and alerts. To overcome this issue, this paper presents the artificial intelligence (AI) and blockchain based proposed architecture on a 6G network. The proposed architecture is examined with a car hacking dataset, wherein the sensors of vehicles are communicating with each other for data sharing. For that, we have adopted an AI algorithm, i.e., random forest (RF), to classify normal and malicious data traffic. Further, edge nodes are considered to reduce the computation of AI algorithms and faster accessibility of vehicular data. Furthermore, incorporating inter planetary file system (IPFS) and a 6G network makes the proposed architecture cost-effective and scalable. Finally, the architecture is evaluated against performance metrics such as accuracy, latency, and scalability. The results demonstrate that the RF surpasses the other algorithms in terms of accuracy and achieves 97% accuracy.

Deep learning and Blockchain-based Essential and Parkinson Tremor Classification Scheme

Sudeep Tanwar (Institute of Technology Nirma University Ahmedabad Gujarat, India); Rajesh Gupta (Institute of Technology, Nirma University, India); Jigna Hathaliya (Nirma University, Ahmedabad, Gujarat, India); Hetav Modi (Institute of Technology, Nirma University, India)

The medical symptoms of Essential tremor (ET) and Parkinson's tremor (PST) are equivalent, including despair, gait, strain, anxiety, and muscular stiffness. In both movement based disorders, neurologists diagnose patients related to clinical evaluations during such hospital visits, so there seems to be a risk of misdiagnosis. To address this issue, Machine Learning (ML) algorithms are being used to identify the patients correctly and distinguish the ET and PST using human-based feature extraction. Motivated by this, we applied Deep Learning (DL) algorithm to overcome the ML issue via automating feature extraction through the model itself. In this paper, we have used the integration of Gated recurrent unit (GRU) and Long short term memory (LSTM) algorithms to predict tremor severity. A recurrent neural network (GRU and LSTM) is a type of neural network that is used to learn temporal connections in time series data. Initially, accelerometer sensors are used to record tremors in all three axial dimensions for each subject. Further, this data is pre-processed using the standard scalar function and scaled in unit variance. Further, this data first passed through the GRU model, and later it fed into the LSTM model to improve the performance of the training and testing model. In addition, the model is evaluated using a blockchain (BC) network, in which an authorized researcher can validate the model, and give the recommendation to improve the model's performance. we have used a smart contract to validate the researcher. The proposed model outperforms with 80.4% training accuracy and 74.1% testing accuracy. The integration of BC and DL makes a system more reliable, transparent, and accurate.

Super Resolution for Augmented Reality Applications

Vladislav Li (Kingston University, United Kingdom (Great Britain)); George Amponis (International Hellenic University, Bulgaria & K3Y Ltd., Bulgaria); Jean-Christophe Nebel and Vasilis Argyriou (Kingston University, United Kingdom (Great Britain)); Thomas Lagkas (International Hellenic University, Kavala Campus & South-East European Research Centre, Greece); Savvas Ouzounidis (K3Y Ltd, Bulgaria); Panagiotis Sarigiannidis (University of Western Macedonia, Greece)

Latest developments in machine learning (ML), adversarial networks, combined with increasingly powerful IoT devices via the introduction of efficient processors, are bringing about the implementation of near real-time object detection and classification for augmented reality (AR) and virtual reality (VR) applications. This paper intends to explore new object detection and classification technologies leveraging super-resolution (SR), that have the potential to be integrated into small, mobile and low-power AR/VR devices. SR in conjunction with novel object detection and classification algorithms are examined in the presented paper, with the ultimate goal of proposing a low-footprint Generative Adversarial Network (GAN)-based framework capable of receiving an LR input and outputting an SR-supported recognition model based on FRRCNN, YOLOv3 or Retina.

Session A4E-PS

Panel Discussion Session

3:00 PM — 4:00 PM EDT
May 2 Mon, 3:00 PM — 4:00 PM EDT

The Emerging Role of AI/ML in Next-Generation Networks

Panelists: Bhavani Thuraisingham (University of Texas, Dallas), Roch Guerin (Washington University, St. Louis), Shiwen Mao (Auburn University)

This talk does not have an abstract.

Session A4E-TS3

Technical Session III

4:00 PM — 5:00 PM EDT
May 2 Mon, 4:00 PM — 5:00 PM EDT

Unsupervised Bias Evaluation of DNNs in non-IID Federated Learning Through Latent micro-Manifolds

Ilias Siniosoglou (University of Western Macedonia, Greece); Vasilis Argyriou (Kingston University, United Kingdom (Great Britain)); Thomas Lagkas (International Hellenic University, Kavala Campus & South-East European Research Centre, Greece); Ioannis Moscholios (University of Peloponnese, Greece); George Fragulis and Panagiotis Sarigiannidis (University of Western Macedonia, Greece)

Lately, Federated Learning (FL) is rapidly evolving in the field of health and quality of life systems. Its ability to train Machine Learning (ML) and Deep Learning (DL) models for the wide variety of Computer Vision (CV) fields that utilize them, one of them being the medical field, while at the same time fortifying the privacy of the sensitive information that describes them, makes the FL technology a necessary tool in modern health and medical CV systems. One of its disadvantages, though, has proven to be the quality of models used for the decentralized learning process and the ability to understand them. Low quality, unethical or biased models used for FL training, usually due to Non-IID (Independent and Identically Distributed) data, can have catastrophic consequences, especially in critical infrastructure in the medical field. In this paper, we tackle the problem of unfairness of DL models in the FL environment by leveraging the ability of latent mapping and representation learning in decision and augmentative DL models while striving to visualize their knowledge distribution. In particular, micro-Manifolds produced from the discovered latent deformations in a DL model are analyzed and, through a proposed quantization pipeline, the fairness of that model is measured in a summary of quantitative metrics. This methodology follows a fully unsupervised model and data agnostic manner to perform ethical evaluation, while it is documented with both medical and widely used benchmark data and DL architectures.

A Hybrid RF-FSO Offloading Scheme for Autonomous Industrial Internet of Things

Dimitrios Pliatsios (University of Western Macedonia, Greece); Thomas Lagkas (International Hellenic University, Kavala Campus & South-East European Research Centre, Greece); Vasilis Argyriou (Kingston University, United Kingdom (Great Britain)); Antonios Sarigiannidis (Sidroco, United Kingdom (Great Britain)); Dimitrios Margounakis and Theocharis Saoulidis (Sidroco Holdings Ltd, Cyprus); Panagiotis Sarigiannidis (University of Western Macedonia, Greece)

The ever increasing demand for bandwidth triggered by data-intensive applications is imposing a considerable burden on the radio-frequency (RF) spectrum. A promising
solution to address the spectrum congestion problem is the adoption of free-space optical (FSO) communications. In this work, we consider a hybrid RF-FSO system that enables the task offloading process from Industrial Internet-of-Things devices
to a multi-access edge computing (MEC)-enabled base station (BS). We propose a solution that minimizes the total energy consumption of the system by deciding whether the RF or FSO link will be used for the task offloading and optimally allocating the device transmission power while taking into account the task requirements in terms of delay. The proposed solution is based on the integer linear programming (ILP) and Lagrange dual decomposition methods. Finally, we carry out system-level Monte Carlo simulations to evaluate the performance of the solution.

Machine learning model for IoT-Edge device based Water Quality Monitoring

Yogendra Kumar (National Institute of Technology Hamirpur, India); Siba Kumar Udgata (University of Hyderabad, India)

The aim of this work is to intelligently detect alarming events in the water quality using machine learning techniques at the edge device which is adaptive to localities, applications and also time. There are four objectives of this work; (1) To develop an edge device for sensing the water quality parameters (2) to detect changes in the water quality with respect to base line parameter using machine learning approach at the edge device itself (3) to generate the alarm signals when water quality parameters goes beyond its threshold value and (3) to classify different types of contamination and analyze them for identifying possible contamination types. For the experimentation, three water quality indicative methods are used to calculate the water quality namely (a) Weighted Arithmetic Index (b) NSF Water Quality Index and (c) User feedback of the water quality. Water quality is determined using water quality indexes (WQI) on the basis of six physico-chemical sensor parameter like biological oxygen demand, dissolved oxygen, pH, total hardness, total dissolved solid and turbidity. With the help of WQI of these methods, a light weight machine learning model which is suitable for the edge device, has been developed using Support Vector Machine (SVM) algorithm. We also clustered the alarming events to find out different types of alarming events.

Session A4E-TS4

Technical Session IV

5:30 PM — 6:50 PM EDT
May 2 Mon, 5:30 PM — 6:50 PM EDT

L3Fog: Fog Node Selection and Task Offloading Framework for Mobile IoT

Mehbub Alam (Indian Institute of Information Technology, Guwahati, India); Nurzaman Ahmed (Indian Institute of Technology, Kharagpur, India); Rakesh Matam (Indian Institute of Information Technology Guwahati, India); Ferdous Barbhuiya (Indian Institute of Technology Guwahati, India)

Due to the mobility of end user and resource-constraint nature of fog nodes in Internet of Things (IoT), tasks offloading and migration are essential but challenging. Moreover, fog node selection as per the location of the mobile devices is still an intricate issue. This paper proposes L3Fog, a learning-based location-aware low-latency fog selection and offloading scheme for IoT. The proposed solution predicts the location of the mobile nodes using machine learning-based approaches. The found location is mapped with the service area of fog nodes. A task offloading decision is taken by using the mapping function to offload to the nearest and suitable fog node. Our approach considers resource and QoS for choosing a fog node for seamless computation. We calculate the location of the mobile device using a real dataset. The proposed scheme's performance analysis shows significant improvement compared to baseline algorithms.

Quality-Aware Incentive Mechanism Design Based on Matching Game for Hierarchical Federated Learning

Hui Du (Beijing Information Sci&Tech University, China); Zhuo Li and Xin Chen (Beijing Information Science & Technology University, China)

To protect user privacy and Combined with mobile edge computing, hierarchical federated learning (HFL) is proposed. In HFL, since participating in model training needs to consume its own resources, it is necessary to design an effective incentive mechanism to incentive mobile devices for training a high quality model. To select high-quality mobile devices to participate in model training, this paper proposes a model quality maximization mechanism MaxQ based on matching game. in MaxQ, the allocation of mobile devices to each edge server is realized so that the sum of the local model quality is maximized. Finally, through a large number of simulation experiments, compared with FAIR and EHFL, the model quality of MaxQ is improved by 10.8% and 12.2%, respectively.

MbRE IDS: An AI and Edge Computing Empowered Framework for Securing Intelligent Transportation Systems

Varun Kohli (National University of Singapore, Singapore); Amit Chougule (BITS Pilani, India); Vinay Chamola (BITS-Pilani, India); F. Richard Yu (Carleton University, Canada)

Recent years have seen a widespread growth of research in the Internet of Things (IoT). While mobility networks such as the Intelligent Transportation Systems (ITS) are being increasingly studied for their application in smart cities, there are numerous cyber threats that may disrupt the security and safety of the users of such networks. This study proposes an intelligent, statistical Intrusion Detection System (IDS) called Multi-branch Reconstruction Error (MbRE) for the long term security of ITS againt known and unknown threats. The proposed IDS learns only from normal behavior, detects deviation of vehicular from it, and classifies it into eight generalized buckets based on the aspects of the data found to be malicious, i.e. frequency, identity and motion (speed and position). The results obtained show the success of the proposed IDS in detecting different threats with recall and accuracy scores between 97.5% to 100% without the need to train on them.

Resource Aware Fog Based Remote Health Monitoring System

Dilwar Hussain Barbhuiya, Adittya Dey, Rajdeep Ghosh, Kunal Das, Kumarjit Ray and Nabajyoti Medhi (Tezpur University, India)

In today's world of medical science, remote patient monitoring devices are becoming more important and a future needs particularly in the present COVID-19 situation as individuals are preferred to be kept isolated. Patients would be benefited from a suitable monitoring system that measures their important medical parameters such as pulse rate, oxygen saturation or SpO2, body temperature, blood pressure, and GSR. This system can increase the medical staff efficiency by drastically decreasing their duties in hospitals and the need to attend to them individually. Patients in their home isolation may utilize the device as well, and their vital indicators may be checked by doctors remotely. In this work, we are prototyping a power-efficient, wearable medical kit and a resource-aware fog network set up to handle IoT data traffic. The idea behind the design is to process the critical medical sensors' data in the fog nodes which are deployed at the edge of the network. The data thus received, is used for a machine learning-based solution for personal health anomalies and COVID infection risk analysis.

Session A4E-CS

Closing Session

6:50 PM — 7:00 PM EDT
May 2 Mon, 6:50 PM — 7:00 PM EDT

Made with in Toronto · Privacy Policy · INFOCOM 2020 · INFOCOM 2021 · © 2022 Duetone Corp.