Large Language Models (LLMs) offer significant potential for enhancing smart home assistants through more natural and responsive automation. However, solely relying on cloud-based LLMs raises concerns about network dependency and user privacy. In this paper, we conduct a comprehensive evaluation to assess the feasibility of deploying LLMs on low-cost devices, enabling smart home automation in terms of latency and energy consumption. In particular, we first build a comprehensive and reproducible testbed that integrates benchmark and well-trained LLMs with the Home Assistant platform on low-cost edge devices, such as Raspberry Pi 4 and Pi 5. Leveraging this testbed, we evaluate on-device LLM inference performance in terms of inference latency, energy consumption, and thermal characteristics, and provide theoretical estimates of remaining runtime in battery-powered settings. We evaluate the on-device performance on two benchmark quantized models (e.g., Home-1B-v3 and Home-3B-v3) for smart homes. Experimental results show that the Raspberry Pi 5 significantly outperforms the Raspberry Pi 4 in terms of low latency and high energy efficiency across both models, due to its superior processing capabilities. In particular, for the Home-1B-v3 model, the mean inference energy decreases from 264.3 J to 133.9 J, representing a 49.3% reduction in consumption, and the mean latency decreases from 40.3 s to 18.7 s, resulting in a 53.4% improvement in delay in Pi 5 compared to Pi 4. The findings of the paper provide valuable empirical insights for energy and latency-aware smart home automation using quantized LLMs on low-cost edge hardware. This work lays the groundwork for future research on the energy efficiency of LLM applications.

Federated Edge Learning (FEL) enables privacy-preserving, on-device training across heterogeneous edge devices, reducing data transfer costs. However, most FEL approaches remain simulation-based and fail to address realistic on-device training asynchrony caused by variations in computing power, data volume, and energy availability among devices. To address the issue, we proposeSPFERE, aSemi-synchronousPower-aware andFairnEss-RegulatedEngine in this paper, designed for power-constrained edge environments and implemented on a real-world edge testbed to support asynchronous model updating, power management, and fairness-aware model aggregation. Specifically, we propose a client grouping-based semi-synchronous aggregation protocol that reduces idle waiting time for power-abundant devices and mitigates stale updates from power-constrained devices, along with our in-depth convergence analysis. Then, we introduce a long short-term memory (LSTM)-based power estimation approach to predict remaining battery voltage for devices with limited communication overhead, enabling early warnings for power dropouts. Lastly, we design fusion-based fairness-aware model aggregation methods to prevent bias by considering device participation frequency and training workload. We systematically validate our framework through experiments on both a simulation platform and a real-device testbed. Our extensive experimental results demonstrate the effectiveness and resilience ofSPFEREin dynamic and heterogeneous edge environments.

In recent years, with the rise of large language models, model sizes have grown dramatically, garnering attention for their remarkable performance but also raising concerns about the substantial computational and communication resources they require. This has created significant challenges in fine-tuning or re-training models on devices with limited computing and memory resources. Efficient model compression through low-rank factorization has emerged as a promising solution, offering a way to balance the tradeoff between compression ratio and prediction accuracy. However, existing approaches to low-rank selection often rely on trial-and-error methods to determine the optimal rank, lacking theoretical guidance and incurring high computational costs. Furthermore, these methods typically treat low-rank factorization as a post-training process, resulting in suboptimal compressed models. In this paper, we design a novel approach by integrating rank selection into the low-rank training process and performing independent layer-wise rank selection under the guidance of a theoretical loss error bound. Specifically, we first conduct a comprehensive theoretical analysis to quantify how low-rank approximations impact the training losses. Building on these insights, we develop an efficient layer-wise rank search algorithm and seamlessly incorporate it into low-rank singular value decomposition (SVD) training. Our evaluation results on benchmark datasets demonstrate that our approach can achieve high prediction accuracy while delivering significant compression performance. Furthermore, our solution is generic and can be extended to broader learning models.

In WiFi and many other sensing applications, researchers have adopted the approach of representing each activity class with a small portion of data to address user variability across different environments, thereby enabling model generalization. This solution raises yet another challenge - limited training samples for each activity class, particularly for a considerable number of classes to be classified. In this paper, we first introduce a visionary architectural layering architecture that enables the end-to-end lifecycle of multi-sensor federated learning, paving the way for realizing federated AI-based sensing. We identify three challenges, e.g., data heterogeneity, scalability under non-IID conditions, and few-shot learning. According to the architecture, we showcase a lightweight, general workflow that utilizes a federated learning framework to address the few-shot learning challenge across various classifications in WiFi Channel State Information (CSI). We analyze this workflow and identify two key factors that most significantly affect its performance (i.e., the aggregation method and the model architecture). Based on the analysis, we propose three network models for CSI-based activity recognition: CSI-AlexNet, CSI-ActNet, and CSI-ResNet. Through extensive performance evaluations, our experimental results show that the proposed approach achieves an accuracy of 97.97% on CSI-ActNet while maintaining low computational demands, making it well-suited for real-time fine-tuning on edge devices with constrained resources.

The restaurant industry struggles with managing high volumes of phone-based customer interactions, particularly during peak hours, leading to operational inefficiencies, increased labor costs, and diminished customer satisfaction. Inspired by the rise of large language models (LLMs) in recent years, in this paper, we introduce an AI-driven customer service application leveraging fine-tuned LLMs to address these challenges by integrating speech recognition, text generation, and text-to-speech technologies for seamless real-time interactions. Using domain-specific conversational data from a local restaurant, the system fine-tunes state-of-the-art LLMs, including Llama-3.2-1B and Llama-2-7B, to generate human-like responses tailored to the restaurant industry. The architecture consists of client-side voice-to-text and text-to-voice components alongside a server-side AI backend, ensuring efficient processing and a smooth user experience. Experimental results show that the fine-tuned Llama-2-7B model delivers superior accuracy and robustness, while GPU utilization significantly enhances response latency and overall performance. Through our designed platform, we demonstrate the potential of domain-specific LLM fine-tuning in transforming customer service applications by automating phone-based interactions, reducing operational costs, and improving customer satisfaction. Our demo video can be reached with the link: https://www.youtube.com/watch?v=K_qQaueAmtA.

Recent years have witnessed the deep involvement of hand gesture recognition in Internet of Things (IoT) devices in healthcare, autonomous driving, virtual reality, augmented reality, etc., since hand gestures offer a natural and intuitive way for human beings to interact with IoT devices without relying on traditional input methods such as keyboards, mice, or touchscreens. Thus, accurate gesture recognition is crucial to its development. Consequently, many recognition approaches are proposed, from the traditional computer vision domain to the deep learning domain. Although with promising recognition performance, these approaches typically come with high computational and energy costs relying on graphics processing unit (GPU) accelerations, which cannot be compatible with low-cost and non-GPU IoT edge devices. To this end, in this paper, we develop an artificial intelligence (AI)-enabled system for real-time hand gesture recognition on low-cost edge devices, e.g., Raspberry Pis. Particularly, we first design a simple but effective convolutional neural network (CNN)-based model with residual blocks to handle American Sign Language (ASL) digits tasks, which enables fast-bust-accurate real-time inferences. Then, to fit the low-cost environment for edge devices, we involve model quantization techniques to shrink the model size, thus requiring reduced memory and storage for edge devices. In addition, we plug a motion sensor into our system to automatically turn off our recognition once it detects no people nearby, reducing energy consumption. By evaluating the performance of real-time hand gesture recognition, our system maintains a high prediction accuracy rate, e.g., over 96 %. Moreover, our designed solution has the potential to be privileged to other common low-cost edge devices.

Object detection is significant in real-world applications, including self-driving cars, surveillance systems, and visionenabled robotic systems, among others. Despite the success of benchmark deep learning-based approaches, like YOLO, achieving high detection accuracy, they are typically computationally intensive and require GPUs to achieve optimal performance, preventing them from being widely deployed on low-power enduser devices. Particularly, when deploying these models on non-GPU devices, their inference speed is significantly degraded due to the lack of GPU support. To this end, in this paper, we propose an optimized object detection model called QATFP-YOLO (Quantization-Aware Training and Filter Pruning on YOLO), aiming to enhance inference speed on non-GPU devices, which could be trained and inferred on local end-user devices without GPU support. To reduce the computation complexity, we propose two optimized training strategies based on our QATFP-YOLO model by considering: (i) model quantization technique that reduces the model size and memory usage without sacrificing accuracy and (ii) filter pruning technique that removes redundant parameters from the model, further reducing memory usage and inference time. By evaluating the performance on a real smartphone, we find our QATFP-YOLO model achieves exceptional inference speeds, reaching approximately 88 frames per second, notably surpassing traditional YOLO-Lite models by over fourfold.

Internet of Things (IoT) devices generate massive amounts of data from local devices, making Federated Learning (FL) a viable distributed machine learning paradigm to learn a global model while keeping private data locally in various IoT systems. However, recent studies show that FL’s decentralized nature makes it susceptible to backdoor attacks. Existing defenses like robust aggregation defenses have reduced attack success rates by identifying significant statistical differences between normal and backdoored models individually. However, these defenses fail to consider the potential collusion among attackers to bypass statistical measures utilized in defenses. In this paper, we propose a novel attack approach, called collusive backdoor attacks (CBA), which bypasses robust aggregation defense by considering both local backdoor training and post-training model manipulations among collusive attackers. Particularly, we introduce a non-trivial perturbation estimation scheme to add manipulations over model update vectors after local backdoor training and use the Gram-Schmidt process to speed up the estimation process. This makes the magnitude of the perturbed poisoned model to the same level as normal models, evading robust aggregation-based defense while maintaining attack efficacy. After that, we provide a pilot study to verify the feasibility of our perturbation estimation scheme, followed by its convergence analysis. By evaluating the attack performance on four representative datasets, our CBA approach maintains high attack success rates under benchmark robust aggregation defenses in both IID (independent and identically distributed) and non-IID local data settings. Particularly, it increases the attack success rate by 126% on average compared to individual backdoor attacks.

With the proliferation of computationally powerful edge devices, edge computing has been widely adopted for wide-ranging computational tasks. Among these, edge artificial intelligence (AI) has become a new trend, allowing local devices to work cooperatively and build deep learning models. Federated learning is one of the representative frameworks in distributed machine learning paradigms. However, there are several major concerns with existing federated learning paradigms. Existing distributed frameworks rely on a central server to coordinate the computing process, where such a central node may raise security concerns. Federated learning also relies on several assumptions/requirements, e.g., the independent and identically distributed (i.i.d.) data and model homogeneity. Since more and more edge devices are able to train lightweight models with local data, such models are normally heterogeneous. To tackle these challenges, in this article, we develop a blockchain-empowered federated learning framework that enables learning in a fully decentralized manner while taking the model heterogeneity and data heterogeneity into account. In particular, a federated learning framework with a heterogeneous calibration process, i.e., Model and Feature Calibration (FL-MFC), is developed to enable collaboration among heterogeneous models. We further design a two-level mining process using blockchain to enable the secure decentralized learning process. Experimental results show that our proposed system achieves effective learning performance under a fully heterogeneous environment.

Moving object trajectory tracking plays an important role in traffic scheduling, route planning, advertising recommendations, and other associated social services. The success of moving object trajectory tracking can be attributed to the extensive use of Internet of Things (IoT) devices, which collect a growing volume of spatio-temporal data. To mine the spatio-temporal correlation, traditionally, recurrent neural networks (RNNs) and their variants, such as long-short-term memory (LSTM) and bidirectional long-short-term memory (BiLSTM), have shown their effectiveness in forecasting moving object positions. However, these methods have faced challenges in dealing with complex temporal dependencies due to the limited memory of storing past information using basic hidden layers. To address this issue, in this study, we propose a spatiotemporal attention-based transformer model to mine the spatiotemporal correlation of moving object trajectories in smart transportation systems, which offers improved performance in long-term trajectory prediction tasks. Moreover, most existing works overlook the importance of sampling issues in the trajectory prediction and tracking process. To this end, we develop an adaptive approach by leveraging spatio-temporal sampling to optimize trajectory tracking with reduced data transmission rates and computational costs. The experimental results on real-world datasets demonstrate the superiority of our transformer-based approach over existing RNN-based methods in trajectory predictions and confirm the feasibility of our optimal sampling solution in enhancing trajectory tracking performance.

The Internet of Things (IoT) paradigm is gradually becoming more prevalent through numerous devices and technologies, including sensors, actuators, microcontrollers, cloud-enabled services, and analytics. IoT objects gain intelligence by integrating with wireless sensor networks (WSNs), mobile computing and communication, and others. With sensors, smart things can be enabled by monitoring and identifying environmental changes related to motion, temperature, humidity, pressure, light, vibration, etc. To timely keep track of state changes, researchers are considering developing a cyber replicator, denoted as Digital Twin (DT), of real physical systems as a way to visualize, model, and work with complex cyber-physical systems (CPS). In this paper, we first refine the dataset to a format that can be easily used for deep learning (DL) experiments, IoT data pipeline development, data modeling and simulation, data aggregation, etc. We then demonstrate that DT data can be used to determine space occupancy based on the ambient light sensor, which tends to indicate occupancy in particular spaces because the building has smart lighting that will switch off when rooms are unoccupied after a certain time. Given the apparent developments in machine learning technology, it is clear that machine learning-based prediction has the ability to enhance resource utilization and further forecast future events. Particularly, we use a DT-based dataset and Long-Short-Term Memory (LSTM) neural network architecture to forecast the campus building’s internal temperature.

With the advance in artificial intelligence and high-dimensional data analysis, federated learning (FL) has emerged to allow distributed data providers to collaboratively learn without direct access to local sensitive data. However, limiting access to individual provider’s data inevitably incurs security issues. For instance, backdoor attacks, one of the most popular data poisoning attacks in FL, severely threaten the integrity and utility of the FL system. In particular, backdoor attacks launched by multiple collusive attackers, i.e., distributed backdoor attacks, can achieve high attack success rates and are hard to detect. Existing defensive approaches, like model inspection or model sanitization, often require to access a portion of local training data, which renders them inapplicable to the FL scenarios. Recently, the norm clipping approach is developed to effectively defend against distributed backdoor attacks in FL, which does not rely on local training data. However, we discover that adversaries can still bypass this defense scheme through robust training due to its unchanged norm clipping threshold. In this paper, we propose a novel defense scheme to resist distributed backdoor attacks in FL. Particularly, we first identify that the main reason for the failure of the norm clipping scheme is its fixed threshold in the training process, which cannot capture the dynamic nature of benign local updates during the global model’s convergence. Motivated by it, we devise a novel defense mechanism to dynamically adjust the norm clipping threshold of local updates. Moreover, we provide the convergence analysis of our defense scheme. By evaluating it on four non-IID public datasets, we observe that our defense scheme effectively can resist distributed backdoor attacks and ensure the global model’s convergence. Noticeably, our scheme reduces the attack success rates by 84.23% on average compared with existing defense schemes.

Internet connected mobile devices in 5G and beyond (simply 5G+) systems are penetrating all aspects of people's daily life, transforming the way we conduct business and live. However, this rising trend has also posed unprecedented traffic burden on existing telecommunication infrastructure including cellular systems, consistently causing network congestion. Although additional spectrum resources have been allocated, exponentially increasing traffic tends to always outpace the added capacity. In order to increase the data rate and reduce the latency, 5G+ systems have heavily relied on hyperdensification and higher frequency bands, resulting in dramatically increased interference temperature, and consequently significantly more weak signals (i.e., signals with low Signal-to-Noise-plus-Interference (SINR) ratio). With traditional detection mechanisms, a large number of weak signals will not be detected, and hence be wasted, leading to poor throughput in 5G+ systems.

The coronavirus disease 2019 (COVID-19) has rapidly become a significant public health emergency all over the world since it was first identified in Wuhan, China, in December 2019. Until today, massive disease-related data have been collected, both manually and through the Internet of Medical Things (IoMT), which can be potentially used to analyze the spread of the disease. On the other hand, with the help of IoMT, the analysis results of the current status of COVID-19 can be delivered to people in real time to enable situational awareness, which may help mitigate the disease spread in communities. However, current accessible data on COVID-19 are mostly at a macrolevel, such as for each state, county, or metropolitan area. For fine-grained areas, such as for each city, community, or geographical coordinate, COVID-19 data are usually not available, which prevents us from obtaining information on the disease spread in closer neighborhoods around us. To address this problem, in this article, we propose a two-level risk assessment system. In particular, we define a "risk index." Then, we develop a risk assessment model, called MK-DNN, by taking advantage of the multikernel density estimation (MKDE) and deep neural network (DNN). We train MK-DNN at the macrolevel (for each metro area), which subsequently enables us to obtain the risk indices at the microlevel (for each geographic coordinate). Moreover, a heuristic validation method is further designed to help validate the obtained microlevel risk indices. Simulations conducted on real-world data demonstrate the accuracy and validity of our proposed risk assessment system.

The Internet of Things (IoT) integrates information collected from multisources and is able to support various intelligent smart city applications, such as industrial manufacturing, power systems, and mobile healthcare. In the big data era, multisourced data are collected on a daily basis, whereas a large part of the data may be irrelevant, redundant, noisy, or even malicious from a machine learning perspective. Feature selection has been a powerful data cleaning technique to reduce data redundancy and improve system performance in machine learning. Inspired by reinforcement learning that learns from its experience, in this article, we propose a novel efficient deep$Q$-network (DQN)-based feature selection method for multisourced data cleaning. In particular, we model the feature selection problem as a competition between an agent and the environment in dynamic states, which is solved by a DQN. Traditional DQN suffers from high computational complexity and requires a significant amount of time in order to converge in the training process. To tackle these challenges, we develop a space searching algorithm called SS to speed up the training process of the DQN agent. To validate the efficacy and efficiency of the proposed method, we conduct extensive experiments on various types of IoT data. Simulation results show that the proposed DQN-based feature selection algorithms achieve much better performance compared with state-of-the-art methods, and are robust under data poisoning attacks.

While traffic modeling and prediction are at the heart of providing high-quality telecommunication services in cellular networks and attract much attention, they have been approved as an extremely challenging task. Due to the diverse network demand of Internet-based apps, the cellular traffic from an individual user can have a wide dynamic range. Most existing methods, on the other hand, model traffic patterns as probabilistic distributions or stochastic processes and impose stringent assumptions over these models. Such assumptions may be beneficial at providing closed-form formula in evaluating prediction performances, but fall short for practice use. In this paper we propose STEP, aspatio-temporal fine-granular user trafficprediction mechanism for cellular networks. A deep graph convolution network, called GCGRN, is constructed. It is a novel combination of the graph convolution network (GCN) and gated recurrent units (GRU), which exploits graph neural network to learn an efficient spatio-temporal model from a user’s massive dataset for traffic prediction. The prototype of STEP has been implemented. Extensive experimental results demonstrate that our model outperforms the state-of-the-art time-series based approaches. Besides, STEP merely incurs mild energy consumption, communication overhead and system resource occupancy to mobile devices. Moreover, NS-3 based simulations validate the efficacy of STEP in reducing session dropping ratio in cellular networks.

Edge computing, an emerging computing paradigm pushing data computing and storing to network edges, enables many applications that require high computing complexity, scalability, and security. In the big data era, one of the most critical applications is multiparty learning or federated learning, which allows different parties to collaborate with each other to obtain better learning models without sharing their own data. However, there are several main concerns about the current multiparty learning systems. First, most existing systems are distributed and need a central server to coordinate the learning process. However, such a central server can easily become a single point of failure and may not be trustworthy. Second, although quite a few schemes have been proposed to study Byzantine attacks, a very common and challenging kind of attack in distributed systems, they generally consider the scenario of learning a global model. However, in fact, all parties in multiparty learning usually have their own local models. The learning methods and security issues, in this case, are not fully explored. In this article, we propose a novel blockchain-empowered decentralized secure multiparty learning system with heterogeneous local models called BEMA. Particularly, we consider two types of Byzantine attacks, and carefully design “off-chain sample mining” and “on-chain mining ” schemes to protect the security of the proposed system. We theoretically prove the system performance bound and resilience under Byzantine attacks. The simulation results show that the proposed system obtains comparable performance with that of conventional distributed systems, and bounded performance in the case of Byzantine attacks.

Community detection is an effective approach to unveil relationships among individuals in online social networks. In the literature, quite a few algorithms have been proposed to conduct community detection by exploiting the topology of social networks and the attributes of social actors. In practice, community detection is usually conducted by third parties, such as advertisement companies and hospitals, with access to social networks for different purposes, which can easily lead to a privacy breach. In this paper, we investigate community detection in social networks aiming to protect the privacy of both the network topology and the users' attributes. We show that with additional prior knowledge, community detection can be performed by querying the information of only a fraction of instead of the entire population. In particular, we first propose a new scheme called differentially private community detection (DPCD). DPCD detects communities in social networks via a probabilistic generative model, which can be decomposed into subproblems solved by individual users. The private social relationships and attributes of each user are protected by objective perturbation with differential privacy guarantees. Then, we propose a parsimonious node affiliation recovery (NAR) algorithm, which is also differentially private, to unveil the community affiliation information of the whole population based on that of the limited number of queried individuals by solving a sparse optimization problem. Through both theoretical analysis and experimental validation using synthetic and real-world social networks, we demonstrate that the proposed DPCD scheme detects social communities under the modest privacy budget. In addition, we show the effectiveness of NAR to perform community detection by querying a limited number of individuals in social networks.

Community detection is an effective approach to unveil social dynamics among individuals in social networks. In the literature, quite a few algorithms have been proposed to conduct community detection by exploiting the topology of social networks and the attributes of social actors. In practice, community detection is usually conducted by third parties like advertisement companies, hospitals, with access to social networks for different purposes, which can easily lead to privacy breaches. In this paper, we investigate community detection in social networks aiming to protect the privacy of both the network topologies and the users’ attributes. In particular, we propose a new scheme called differentially private community detection (DPCD). DPCD detects communities in social networks via a probabilistic generative model, which can be decomposed into subproblems solved by individual users. The private social relationships and attributes of each user are protected by objective perturbation with differential privacy guarantees. Through both theoretical analysis and experimental validation using synthetic and real world social networks, we demonstrate that the proposed DPCD scheme detects social communities under modest privacy budget.

The ever-growing proliferation of mobile devices equipped with accelerometers has provided new opportunities to capture the semantic meanings of human activities and improve user experience with behavior-based recommendations, which heavily rely on the accuracy of the recognition of daily human activities. Acceleration-based human activity recognition (HAR) is a challenging problem because each accelerometer records multi-dimensional signals in both spatial and temporal domains that have different attributes for representing different activities or even the same activity. Thus we cannot directly compare these signals with each other, because they are embedded in a non-metric space. In this paper, we present a Personalized Recurrent Neural Network (PerRNN) to dynamically segment and recognize the human activities based on accelerometer data. Enlightened by the idea of spatiotemporal predictive learning, the proposed architecture is capable of memorizing different acceleration signals' appearances and temporal variations in a unified memory pool. We evaluate the performance of the proposed framework on a commonly used dataset, WISDM. Experiment results show that compared with state-of-the-art schemes, our proposed PerRNN system recognizes 6 different human activities with the highest overall accuracy of 96.44%.

Studies find that deep learning models are vulnerable to deliberate adversarial manipulations by attackers. Adversarial training is an effective approach to address this problem. Previous works quantize the input sample space to find appropriate perturbations on the benign samples so as to generate adversarial samples for adversarial training. However, since only the input sample space is quantized with the perturbation space being still continuous, finding the optimal perturbation noise is still a non-convex and computationally expensive problem. Moreover, in this case, the found perturbation noise that will be used to generate an adversarial sample may be strong in the continuous search space, but may become weak after quantization in the input sample space. In this paper, we first develop an Iterative Quantized Local Search (IQLS) algorithm that finds strong perturbation noises by quantizing both the input space and perturbation space. Then, we theoretically analyze and prove the upper bound on the number of iterations needed for the IQLS algorithm, based on which we devise an efficient and effective Quantized Adversarial Training (QAT) scheme. Experiment results on six public datasets show that our proposed scheme outperforms state-of-the-art methods to defend against different adversarial attacks. Particularly, QAT improves the system performance by 14%, 11%, 16% on average on CIFAR-10, SVHN, and CIFAR-100 datasets respectively compared with the existing defense schemes, and reduces the computing time by about 60%.

Unsupervised anomaly detection on multidimensional time series data is a very important problem due to its wide applications in many systems such as cyber-physical systems, the Internet of Things. Some existing works use traditional variational autoencoder (VAE) for anomaly detection. They generally assume a single-modal Gaussian distribution as prior in the data generative procedure. However, because of the intrinsic multimodality in time series data, previous works cannot effectively learn the complex data distribution, and hence cannot make accurate detections. To tackle this challenge, in this paper, we propose a GRU-based Gaussian Mixture VAE system for anomaly detection, called GGM-VAE. In particular, Gated Recurrent Unit (GRU) cells are employed to discover the correlations among time sequences. Then we use Gaussian Mixture priors in the latent space to characterize multimodal data. The proposed detector reports an anomaly when the reconstruction probability is below a certain threshold. We conduct extensive simulations on real world datasets and find that our proposed scheme outperforms the state-of-the-art anomaly detection schemes and achieves up to 5.7% and 7.2% improvements in accuracy and F1 score, respectively, compared with existing methods.

Paradigm-shifting systems such as cyber-physical systems, collect data of high- or ultrahigh- dimensionality tremendously. Detecting outliers in this type of systems provides indicative understanding in wide-ranging domains such as system health monitoring, information security, etc. Previous dimensionality reduction based outlier detection methods suffer from the incapability of well preserving the critical information in the low-dimensional latent space, mainly because they generally assume an isotropic Gaussian distribution as prior and fail to mine the intrinsic multimodality in high dimensional data. Moreover, most of the schemes decouple the model learning process, resulting in suboptimal performance. To tackle these challenges, in this paper, we propose a unified Unsupervised Gaussian Mixture Variational Autoencoder for outlier detection. Specifically, a variational autoencoder firstly trains a generative distribution and extracts reconstruction based features. Then we adopt a deep brief network to estimate the component mixture probabilities by the latent distribution and extracted features, which is further used by the Gaussian mixture model to estimate sample densities with the Expectation-Maximization (EM) algorithm. The inference model is optimized jointly with the variational autoencoder, the deep brief network, and the Gaussian mixture model. Afterwards, the proposed detector identifies outliers when the estimated sample density exceeds a learned threshold. Extensive simulations on six public benchmark datasets show that the proposed framework outperforms state-of-the-art outlier detection schemes and achieves, on average, 27% improvements in F1 score.