This work proposes a novel MLaaS Performance Drift Detection framework for IoT environments that employs an MLaaS extraction model that learns service behavior from input-output pairs and identifies prediction-influenced features, and designs an Adaptive-Temporal Performance Drift Detection Mechanism that dynamically adjusts monitoring frequency based on behavioral and data variations.
Abstract
Machine Learning as a Service (MLaaS) is a powerful cloud paradigm enabling data-driven intelligent applications in Internet of Things (IoT) environments, widely adopted across healthcare, smart homes, and industry due to its cost-effectiveness. However, the dynamic nature of IoT frequently alters data distributions, affecting MLaaS stability, while periodic MLaaS updates further introduce performance drift. Unlike traditional ML systems, MLaaS clients operate as black-box users without access to internal data or parameters, making drift detection particularly challenging. To address this, we propose a novel MLaaS Performance Drift Detection framework for IoT environments. The framework first employs an MLaaS extraction model that learns service behavior from input-output pairs and identifies prediction-influenced features. Building on this, the proposed MLaaS Performance Drift Detection (MPDD) model jointly captures variations in input data and MLaaS behavior. We further design an Adaptive-Temporal Performance Drift Detection Mechanism (APDDM) that dynamically adjusts monitoring frequency based on behavioral and data variations, enabling timely drift detection for effective service management. Extensive experiments on real-world datasets demonstrate that MPDD achieves up to 22-25% accuracy improvement over baseline drift detection methods. APDDM provides an average accuracy gain of approximately 4% and reduces the miss detection rate by around 9% compared to fixed-interval monitoring.
The ubiquitous use of Internet of Things (IoT) system in dynamic and real-life contexts presents a high level of challenges because of the constant changes in the data distributions, otherwise known as concept drift. The conventional machine learning models that are implemented in IoT systems are usually trained in static mode and cannot respond to changing trends in data, which causes deterioration in performance as time goes by. The drawback restricts the reliability and independence of intelligent IoT applications in the long run. In order to overcome these challenges, this paper will present a omprehensive lifelong learning model that can be used to enable drift-robust and autonomous IoT intelligence. The framework proposed combines the real-time drift detection, incremental model adaptation with memory-based knowledge retention into a single framework. It allows the IoT systems to continuously learn stream data and retain the past knowledge, thus eliminating catastrophic forgetting. The architecture is edge deportable, which means that it is low-latency inference and less reliant on centralized retraining. This is proven by on-the-job evidence of the proposed approach; it has been shown to be able to sustain stable performance even when data distribution changes, it converts to concept drift faster and it has a higher level of robustness than the traditional batch and online approaches to learning. The framework offers a scalable and effective way of facilitating self-adaptive, resilient, and autonomous IoT systems in the new generation smart environments.
T.Muthumanickam, L. S, D. Jayalakshmi et al.· 2026 6th International Confe...· 0 citations
Machine learning models deployed in sensor-based production environments are prone to performance degradation due to evolving data distributions, commonly known as data drift. In gas sensor systems, such drift often arises from environmental variability and sensor aging, which can significantly reduce predictive reliability if left unaddressed. This study presents an integrated Machine Learning Operations (MLOps) framework that combines performance monitoring, distribution-based drift detection, and adaptive retraining within a unified production pipeline. Experiments are conducted using the Gas Sensor Array Drift Dataset, organized into sequential batches to emulate real-world deployment conditions. Data drift is quantified using Population Stability Index (PSI) and Kullback–Leibler Divergence (KL), which serve as decision thresholds for triggering retraining. The proposed adaptive retraining strategy is systematically compared with baseline (no retraining) and periodic retraining approaches. The results indicate that the adaptive strategy maintains more stable performance across data batches while minimizing unnecessary retraining processes. Additionally, the use of containerization and experiment tracking ensures reproducibility and supports full lifecycle traceability. Overall, this study demonstrates that integrating MLOps practices into sensor-based machine learning systems is essential for improving robustness and ensuring long-term operational sustainability in dynamic environments.
The analysis of large amounts of data generated in real time by applications and services of the Internet of Things (IoT) has become a real challenge, since the data streams are continuous and non-stationary, therefore they require a high quality classification and prediction. Therefore, the use of static machine learning models becomes inefficient, since the behavior of the devices and the distribution of the sensors can change over time, and they require an adaptive model that can be updated in real time. In this work, an adaptive machine learning framework for real-time device activity recognition and future-state prediction is proposed. It is tested on a simulated IoT stream with 30,000 observations and 5 different device states (sleep, sensing, processing, transmitting, idle) under different types of concept drift (abrupt, gradual, incremental, and recurring). The framework consists of an online stochastic gradient descent (SGD) classifier, a recent-window random forest, dynamic model weighting, and Page–Hinkley drift detection. A prequential test-then-train approach is used after an initial set of 2,000 observations. For the task of activity recognition, the framework achieves an accuracy of 93.92% and a corresponding macro-F1 score of 93.81%, outperforming a static random forest as well as a non-adaptive online SGD model by 9.39% and 1.83% percentage points, respectively. For the task of one-step-ahead state prediction, it achieves an accuracy of 68.32% and corresponding macro-F1 score of 68.27%. Average processing latency per observation is below 0.5 ms. The results of this work show that adaptive online learning can efficiently support reliable IoT classification and prediction in changing operational scenarios, while keeping low computational latency.
A. S. Naik, R. E. Franklin Jino, Poonam Singh et al.· International journal of com...· 0 citations
Real-time anomaly detection in industrial IoT (IIoT) often requires processing continuous data streams on resource-constrained edge nodes while addressing non-stationary data distributions caused by changes in device operating states. Traditional statistical or distance-based methods typically rely on fixed thresholds or static models, making it hard to maintain stable performance under complex conditions. Deep learning approaches are computationally intensive, making them unsuitable for resource-constrained edge devices. Existing methods struggle to balance dynamic adaptability to data distributions with computational efficiency. This paper proposes DAStream, a streaming anomaly detection method for edge environments. We propose a dual-phase online clustering framework with statistical enhancement to improve model stability. We design an adaptive anomaly scoring method that uses time-varying statistics to capture dynamic drifts in the data distribution, combined with an online-updated normalized deviation metric. We also propose a lightweight self-calibrating discrimination mechanism to enable dynamic decision boundaries. DAStream relies solely on recursive statistical computations, ensuring constant computational complexity and enabling efficient deployment on devices such as FPGAs. Experiments on typical IIoT datasets and an FPGA platform show that DAStream achieves a $\text{6 2. 6 \%}$ increase in throughput and a 40.7% reduction in energy consumption while maintaining detection performance comparable to the state-of-the-art method, validating its effectiveness and feasibility for IIoT edge intelligence scenarios.
Xiao Liu, Shubo Liu, Zhaohui Cai et al.· Fall Joint Computer Conferen...· 0 citations
The rapid adoption of cloud-based architectures has increased system scalability and flexibility while simultaneously expanding the attack surface and operational complexity of modern applications. Traditional rule-based monitoring systems, which depend on static thresholds and predefined signatures, struggle to detect sophisticated threats and performance irregularities in highly dynamic, elastic, and ephemeral cloud environments where workloads scale up and down continuously and services are frequently redeployed. This paper explores the design and implementation of AI-powered anomaly detection frameworks tailored for cloud-native infrastructures, examining the theoretical foundations, architectural components, and practical deployment considerations of intelligent monitoring systems. By leveraging machine learning techniques such as supervised learning, unsupervised clustering, and deep learning models including recurrent neural networks and autoencoders, artificial intelligence systems can identify deviations from baseline behavior across distributed services, containers, and microservices in real time. The proposed approach integrates telemetry data from logs, metrics, and network traces to establish adaptive behavioral profiles, emphasizing automated feature extraction, continuous model training, and feedback loops that reduce false positives while improving detection accuracy. The framework is explicitly designed to operate across the full lifecycle of anomaly management, from raw data ingestion through model inference to alert generation and remediation. Additionally, this study addresses scalability challenges, data privacy considerations, and integration with DevOps and SecOps workflows. Experimental evaluation, conducted on a dataset exceeding 500,000 records drawn from logs, metrics, and network traffic, demonstrates improved detection rates, faster incident response times, and enhanced system resilience compared to conventional monitoring tools. Five model families were benchmarked side by side, with hybrid ensemble approaches achieving the strongest overall results. The findings suggest that AI-powered anomaly detection significantly strengthens observability and security in cloud-based applications, enabling proactive threat mitigation and operational optimization in increasingly complex distributed environments.
Harsh Verma· International Journal of Sci...· 0 citations
The rapid growth of data from sensors, social media, financial systems, and IoT devices has made streaming data processing a critical research area. Traditional batch learning methods are unsuitable for streaming environments due to memory, time constraints, and inability to handle concept drift. Online learning algorithms provide an effective solution by continuously updating models with incoming data. This paper presents a comprehensive study of online learning techniques for streaming data, focusing on adaptability, accuracy, memory efficiency, and performance. Algorithms such as Stochastic Gradient Descent, Online Passive-Aggressive methods, and online ensemble techniques are analyzed along with their mathematical foundations and trade-offs. The study also addresses challenges like non-stationary data, real-time processing, and scalability, along with solutions such as concept drift detection and adaptive learning. Applications in fraud detection, predictive maintenance, healthcare, and recommendation systems are discussed. Experimental results show that hybrid online ensemble methods outperform traditional single-model approaches in stability and performance, while lightweight algorithms are suitable for edge computing. The paper concludes with future directions including federated learning, reinforcement-based adaptation, and integration with deep learning, emphasizing the importance of online learning in real-time intelligent systems.
Tendai Chikore· International Journal of App...· 0 citations
