This study introduces an advanced federated learning framework tailored for smart home intrusion detection, incorporating knowledge distillation and transfer learning to tackle escalating threats to IoT devices. In light of the rapid expansion of IoT devices and their vulnerability to botnet incursions, our approach specifically addresses the challenges related to the privacy concerns of home device data, heterogeneity of devices, sparse intrusion data, and the dynamic nature of smart home settings. We adaptively select model architectures tailored to the computational capabilities of each device, ranging from simple Neural Networks (NNs) to more complex Convolutional Neural Networks (CNNs) and hybrid CNN-LSTM models, ensuring efficient local training without overburdening the devices. However, it can achieve good performance through collaborative learning, even for devices with lower capacity and sparse data. Our evaluation, conducted using the N-BaIoT dataset, demonstrates the effectiveness of our approach in detecting anomalies across a diverse set of IoT devices infected with real-world botnets such as Mirai and BASHLITE. The results highlight the potential of our framework to provide a robust, privacy-preserving, and adaptable solution for securing smart homes against emerging threats.
Road obstacle detection is one of the crucial tasks for the safe and efficient operation of autonomous vehicles. Existing detection methods often try to detect all objects in the image (i.e. a frame in a video stream). However, it is important to note that not all objects are equally dangerous and that the detection system should not react equally to them. To reduce computation costs (e.g. time and power consumption) in a mobile device, this study introduces a cost-sensitive obstacle detection method to leverage the performance of YOLIC (You Only Look at Interested Cells), a method proposed by us previously. The new method allows YOLIC to pay more attention to important areas for driving. In the experiments, we assign high weights to areas close to the vehicle, and the weights vary depending on the detection distance, direction, and relation to driving. Experimental results on two road obstacle datasets demonstrate that the proposed cost-sensitive detection method can effectively reduce the costs in the most dangerous areas of a given image compared with the baseline YOLIC model. Moreover, our fastest model can achieve real-time performance on a Raspberry Pi 4B, making it possible to deploy the proposed method in low-cost vehicles such as scooters and delivery robots.
The Internet of Medical Things (IoMT) has emerged substantial growth within the healthcare sector, spurred by advancements in smart devices that generate and process critical healthcare data. Sharing this data among trusted healthcare entities is vital for efficient patient care but raises substantial security and privacy concerns. Existing solutions have focused on authentication techniques employing Self-Sovereign Identity (SSI) and Zero-Knowledge Proofs (ZKPs), aiming to ensure secure, auditable, and anonymous authentication. These solutions often prioritize either lightweight authentication or robust security, yet a flexible approach that integrates both characteristics is essential for adaptive cross-domain authentication in the IoMT landscape. Furthermore, the importance of lightweight authentication combined with adaptive verification, pivotal for scaling access control in cross-domain environments, has been underexplored in existing frameworks. Addressing this gap, we proposed a scheme called LSAC which is a lightweight, scalable, and anonymous authentication for SSI-based cross-domain authentication for IoMT setting. Our proposed LSAC scheme not only facilitates rapid and robust identity verification across multiple IoMT domains within a consortium blockchain network but also incorporates the advanced ZK-STARK and Plonk ZKP protocols to enable efficient and secure verification processes. By leveraging Hyperledger for generating smart contracts and managing user interactions across domains, our system represents a significant step forward in cross-domain authentication. Our comprehensive functionality and performance analysis confirms the effectiveness of our solution in providing scalable and secure access to IoMT resources, supporting the ongoing evolution of healthcare services.
A suitable amount of training data is important in efficiently training high-quality deep learning models. In this paper, we performed an empirical study aimed at finding optimal amount of labeled waterfowl annotations or images for training detection and classification deep learning models with real-world aerial images. We proposed a new class-aware sampling method that can generate better training sets which led to significantly better waterfowl detectors. The results show that the trained Faster R-CNN and YOLOv5 models achieved over 86% accuracy with only 300 waterfowl training annotations using the new sampling method while the basic random sampling method needs 500 annotations to achieve same performance. In addition, the class-aware sampling method significantly accelerated training convergence of all detection models we applied. For waterfowl classification, we compared experimental results of two representative supervised-learning models and two semi-supervised learning methods, and combined with two loss functions and two sampling methods to generate training sets. The results show that semi-supervised learning methods outperformed supervised learning models, achieving F1 score over 0.8 using only 100 labeled training images per waterfowl class.
Nowadays, camera traps are widely employed in monitoring biodiversity and assessing the population density of animal species. A challenge in animal recognition in camera trap images is the detection of small animals in complex environments and the identification of heavily obscured animals. This paper presents two novel methods that leverage sequentially captured images to improve animal recognition accuracy: one utilizing optical flow information and the other a motion-based algorithm based on the principle of median filtering. In experiments, we used two new real-world sequence-based camera trap image datasets to evaluate these methods. Our findings indicate that optical flow information effectively reduces false positive cases, while the motion-based algorithm significantly improves the accuracy of detecting animal presence and counting by substantially reducing false negative cases. Specifically, using the MegaDetector with a confidence threshold of 0.5 as the baseline, the motion-based method reduced false negative cases by over 70% while only slightly increasing false positive cases, and improved animal counting accuracy by more than 25%.
Parkinson’s disease (PD) is a progressive neurological disease primarily impacting movement. Deep brain stimulation (DBS) stands out as a potent therapeutic treatment for addressing PD’s motor symptoms. Nevertheless, optimizing DBS parameters to enhance both effectiveness and efficiency remains a significant challenge. Traditional DBS relies on high-frequency stimulation, lacking adaptability to adequately dress the dynamic nature of PD symptoms. In this paper, we develop a novel method to enhance adaptive DBS utilizing reinforcement learning (RL), while concurrently reducing computational costs through neural network quantization. Our approach entails the discretization of state and action spaces, representing the status of neurons within PD-associated brain regions and the corresponding stimulation patterns, respectively. The primary learning objective is to maximize the accumulated treatment reward, specifically reduced power in the beta frequency band, while maintaining a low stimulation frequency. Additionally, to enhance computing and inference efficiency, we integrate two types of neural network quantization approaches, post-training quantization (PTQ) and quantization-aware training (QAT), with the RL model. Experimental evaluations conducted within a computational model of PD demonstrate the efficacy of our RL approach, even after quantization. While quantization induces some alterations in RL model outputs, the overall DBS efficacy remains unaffected. Furthermore, the animal testing of a rat with PD confirms the effects on both animal behavior and neural activity, underscoring the potential of RL-based adaptive DBS as a promising avenue for personalized and optimized treatment of PD.
The promise of artificial intelligence (AI) has prompted various industries to explore how it can enhance their businesses and improve customer experiences. However, the practical deployment of AI services faces challenges due to internal infrastructure limitations, data policies, and network restrictions. In this paper, we present our current project—a solution designed to empower mental health centers with AI tools while adhering to stringent data and network constraints. Our approach involves deploying AI services and data within internal enterprise networks, while the customer-facing mobile app resides in the public cloud. We introduce a custom AI agent system that monitors requests from the public cloud, enforces data access policies, schedules GPU computing resources, and performs model inference. All computations and raw data storage remain confined within the enterprise network. Additionally, we develop a comprehensive multi-modal AI services app, encompassing semantic search, data indexing, document and summarization, question-answering, and translation services for multilingual users. Administrative users have the ability to correct AI-generated data and contribute to continuous model refinement. Our solution serves as a blueprint for enterprises facing similar restrictions.
According to a PRNewswire research report, the drone cloud surveillance intelligence market will witness a YOY growth of 12.05% in 2023, with a CAGR of 13.58% during the forecast period. This burgeoning demand has spurred an increased interest in the research and development of cloud platforms for drone integration services. In this paper, we define the drone cloud concept, propose a surveillance drone cloud with reference infrastructure, and discuss various drone cloud service and deployment models. We also address challenges, issues, and needs associated with surveillance drone AI services and taxonomy. Furthermore, we explore opportunities in integrating drone technology with cloud computing, real-time data sharing, and drone-human interaction support, while proposing a smart drone service model for efficient resource utilization. Our analysis offers insights into the present state and future directions of drone cloud surveillance technology.
Detecting real-time distress signals in public places can help authorities react promptly in life-threatening circumstances or violent encounters and assist those in distress. The most common way to ask for help is to shout or flap your hands above your head, often called a “waving” action. This study aims to deploy distress signal recognition software on existing surveillance systems or cameras to automatically identify risky situations and alert the authorities. This will reduce human interaction and the constant monitoring of numerous surveillance feeds. In this paper, we propose two different methods for distress signal recognition. Both approaches use pose estimation models, which are used to identify and locate skeletal points on the human body. In the first approach, we propose an algorithm that detects a series of features that are used to identify a distress signal action. In the second approach, we propose a neural network-based approach to recognize distress signals. The neural network is trained on a dataset that we have constructed by identifying major features calculated using the skeletal points from the pose estimation model. Our proposed approach uses significantly fewer training data compared to the traditional convolutional neural network (CNN) approaches that use image or video data for action recognition. In this paper, we compare and assess the two approaches and demonstrate their effectiveness on a real-world dataset.
Mobile social networks have the potential to become the most effective platform for delivering urgent information during natural disaster events. Social media applications like Twitter are ideal for real-time news delivery but lack many important features to become an effective emergency notification system. The most significant challenge is determining which news is an emergency and whether it is relevant to the user. In this project, we propose an intelligent emergency notification social network application that can automatically perform data mining from online social media platforms, filter out non-emergency and unrelated news, and deliver only classified emergency notifications to users. Rather than using multiple separate systems or machine learning models to achieve our goals, we propose a joint Multi-class Text Classification model based on BERT to filter out non-emergency tweets and classify the type of emergencies in one end-to-end neural network model. Additionally, we train a Named Entity Recognition (NER) model to extract locations from the classified emergency news. We evaluated our system in multiple historical emergency events and demonstrated the effectiveness of emergency news delivery using various performance metrics.
The ever-increasing demand for reliable data storage solutions at mobile edge computing makes it a necessity to develop timely fault detection mechanisms. In this paper, we address the research challenge of detecting failed data queries in an application and infrastructure agnostic way. In this context, the mobile edge infrastructure includes a Database Management System (DBMS) that runs on server nodes and client nodes that retrieve, store, update and delete data in the DBMS. Specifically, we propose a rule-based algorithm with thresholds that takes as input utilization metrics and detects the faults. It does so without using sensitive DBMS credentials. To verify the applicability and the generality of the proposed algorithm, we made an experimental evaluation with six different query generation functions and testbed configurations. The comparison with other popular machine learning methods used for anomaly detection in monitoring systems, showed that the proposed algorithm significantly surpasses the other methods in terms of Precision and F1-score. During the operation of the mobile edge computing we gathered the utilization metrics, the queries success and fail status and created six datasets that we make them publicly available.
Real-time data processing is a standard requirement in Fog Computing. Dynamically adapting data stream processing frameworks is an essential functionality to handle time-varying workloads efficiently and to optimize resource consumption. However, horizontal scaling alone, by adapting the parallelism and number of provisioned nodes, faces limits when available compute resources are scarce. We propose TransScale, a combined-approach auto-scaler that combines horizontal scaling to approximation computing, controlling it through transprecision computing. We design TransScale to make the approximation method transparent to the system and support context-specific requirements through QoS-driven re-configuration decisions. Based on the policy’s objective, we show that it can reduce re-configuration occurrences, optimize resource utilization and sustain high workloads in resource-constrained environments.
Nowadays there is an increasing interest from the edge computing and IoT community for virtual sensors due to their advantages of low cost, robustness, easy installation and multi-purpose use. Virtual sensors are software components capable of replacing physical sensors providing aggregations and higher representations of physical measurements. Since virtual sensors rely on taking input and process measurements from external sources of data, they bear their limitations. To this end, in this paper, we tackle the challenges of missing values and low sampling rate. Specifically, our research goal is to design a virtual sensor that operates smoothly even if it misses some input values. Additionally, even if the sampling rate of the external input is low, the virtual sensor will be capable of providing output values in a higher rate. In order to achieve these functionalities we examine and tailor different lightweight deep learning models appropriate for an edge computing setting. For the experimental evaluation we also developed an IoT platform and run a smart home use case with humidity and temperature sensors. Comparing the evaluation outcomes of our methodology with baseline missing values techniques and multi-step approaches, our proposed methodology is proved to be promising and accurate.
From the viewpoint of food self-sufficiency, aquaculture is attracting attention. Since aquaculture is a typical labor-intensive industry, it is desirable to improve productivity through the use of IT. Therefore, this paper presents a study of Aqua Colony for fully automated aquaculture. Aqua Colony automates the aquaculture industry, which has relied on manual labor, by utilizing IoT, AI, and drones. Specifically, the AI calculates the optimal amount of food to be fed based on information from the IoT like sensors, and the drone automatically feeds the fish. This paper describes overview of Aqua Colony, and the automatic feeding system, which is characterized by optical flow to a Support Vector Machine (SVM) to determine the degree of fish activity.
One of the critical factors for a successful emergency mission is reliable communication between emergency responders and the community. In today’s world, there is an increasing need to collect and provide meaningful information to communities and their fire service organizations in real-time to assist with informed decision-making and early hazard alerts. Our proposed solution to this problem is a multi-modal emergency communication system based on NB-IoT, C-V2X, and nearby networking technologies. Our system is designed to converge various services, such as data collection from sensors deployed in remote industrial and residential areas, wildfire detection and monitoring, nearby communication for first responders, vehicle-to-vehicle communication for fire trucks, and ground-to-aerial communication for UAVs. We utilize the Container Registry provided by the Azure platform for over-the-air updates and to deploy new modules and updates. In addition to the hardware systems, we propose a cloud dashboard to manage all devices and offer real-time data visualization, device tracking, asset management, and live-video streaming. These features facilitate highly reliable communication among responders with low delay, enabling them to respond quickly and efficiently to emergencies.
In recent years, there has been an increase in cyber attacks in mobile cloud environment. Intrusion Detection Systems (IDS) have played an important role in protecting mobile cloud security. Many techniques have been utilized to implement IDS, among them, machine learning-based techniques have generated promising results. Especially, complex deep neural networks show a higher detection rate than traditional machine learning models. However, the interpretation of the decision made by a neural network becomes harder to understand as its architectural complexity increases. This challenge makes it difficult for the human experts to fine-tune their detection systems, trust the detection system’s results, and make decisions accordingly when IDS systems are deployed. To address this issue, we propose an explainable intrusion detection framework that employs deep learning mechanisms to identify cyber-attacks and utilizes knowledge graph as the knowledge foundation to add human understanding of machine learning and explain the machine learning results. The use case study demonstrates that the proposed framework can not only successfully identify network intrusions but also effectively reveal important information about its internal working mechanisms of the mysterious deep learning Blackbox.
Most cryptographic-based access control schemes implemented in IoT-based and mobile cloud computing generally provide secure and lightweight privacy-preserving data sharing and access for data users. Ciphertext policy attribute-based encryption (CP-ABE) is one of the suitable techniques used to support fine-grained and secure data sharing in data outsourcing environment. However, its cryptographic construct is based on pairing and exponentiation which are expensive operations that are not practical to be run in the resource-constraint devices. In this paper, we propose a secure and efficient mobile-cloud based access control based on the fully outsourced CP-ABE decryption. Essentially, we introduce a transformation key technique to allow a mobile user to compute the secret key upon the decryption. Our proposed scheme outperforms the existing works in the way that there is no secret key retained at the mobile device and no cost of CPABE decryption at the mobile client side. Finally, we conducted performance evaluation to substantiate that our proposed scheme is efficient in practice.
In order to plan rapid response during disasters, first responder agencies often adopt ‘bring your own device’ (BYOD) model with inexpensive mobile edge devices (e.g., drones, robots, tablets) for complex video analytics applications, e.g., 3D reconstruction of a disaster scene. Unlike simpler video applications, widely used Multi-view Stereo (MVS) based 3D reconstruction applications (e.g., openMVG/openMVS) are exceedingly time consuming, especially when run on such computationally constrained mobile edge devices. Additionally, reducing the reconstruction latency of such inherently sequential algorithms is challenging as unintelligent, application-agnostic strategies can drastically degrade the reconstruction (i.e., application outcome) quality making them useless. In this paper, we aim to design a latency optimized MVS algorithm pipeline, with the objective to best balance the end-to-end latency and reconstruction quality by running the pipeline on a collaborative mobile edge environment. The overall optimization approach is two-pronged where: (a) application optimizations introduce datalevel parallelism by splitting the pipeline into high frequency and low frequency reconstruction components and (b) system optimizations incorporate task-level parallelism to the pipelines by running them opportunistically on available resources with online quality control in order to balance both latency and quality. Our evaluation on a hardware testbed using publicly available datasets shows upto $\sim 54$% reduction in latency with negligible loss $(\sim 4 -7$%) in reconstruction quality.
Deep neural networks (DNN) have enabled dramatic advancements in applications such as video analytics, speech recognition, and autonomous navigation. More accurate DNN models typically have higher computational complexity. However, many mobile devices do not have sufficient resources to complete inference tasks using the more accurate DNN models under strict latency requirements. Edge intelligence is a strategy that solves this issue by offloading DNN inference tasks from end devices to more powerful edge servers. Some existing works focus on optimizing the inference task allocation and scheduling on edge servers. Other works focus on dynamically adapting the inference quality. In this work, we propose combining strategies from both research areas to serve applications that use deep neural networks to perform inference on offloaded video frames. The goals of the system are to maximize the accuracy of inference results and the number of requests the edge cluster can serve while meeting latency requirements. We propose heuristic algorithms to jointly adapt model quality and route inference requests, leveraging techniques that include model selection, dynamic batching, and frame resizing. We evaluated the proposed system in testbed experiments, comparing it to a baseline deployment with no quality adaptation. Our system provided 9.7% higher accuracy at low loads and met deadlines for 43.5% more frames at high loads. We also evaluated the proposed system in simulated experiments, where the system processed 16% more accurate frames per second than a solution with only quality adaptation and 45% more than one with only intelligent request routing.
In recent years Cloud Robotics technology has been proposed to overcome the constraints imposed by the resources of standalone robots. We can imagine that in near future robots will be very present in everyday life and interact with humans, so it is necessary to guarantee that robots could make decisions even if the connection to the cloud is unavailable. It is then important to move the critical tasks on the edge devices in order to make them always accessible, not following the Cloud Robotics paradigm but the Dew Robotics one instead. In this paper we propose DewROS, a platform for Dew Robotics that uses monitoring entities to monitor the system status in order to adapt the application operating conditions. In particular in this work we describe the DewROS platform and its application in the case of video analysis in a surveillance scenario. The results provided in this paper demonstrate how DewROS allows us to exploit at their best the limited resources of our robots.
