Empirical Evaluation of Edge Computing Offloading in Smart Homes

Thesis Abstract

The rapid growth of smart homes and the proliferation of latency-sensitive and compute-intensive applications have underscored the need for efficient edge computing offloading strategies that balance performance, energy consumption, and privacy. This study addresses the problem of determining when and how to offload processing tasks from smart-home devices to local edge servers versus the cloud, in order to optimize latency, network utilization, and user experience while preserving user privacy. The aim is to develop an empirical understanding of offloading decisions in heterogeneous smart-home environments and to identify design guidelines for adaptive offloading frameworks. Specific objectives include (1) to quantify the impact of offloading decisions on end-to-end latency, energy consumption, and cost across varying network conditions and device capabilities; (2) to evaluate the accuracy of predictive models that classify tasks by offloadability using features such as task size, computation intensity, sensor data locality, and privacy requirements; (3) to assess user-perceived quality of experience (QoE) under different offloading strategies; (4) to examine the interaction between offloading policies and security/privacy assurances in edge versus cloud configurations; and (5) to propose an adaptive offloading framework guided by empirical findings. The methodology adopts a mixed-methods, empirically driven design anchored in realism and reproducibility. The population comprises smart-home ecosystems incorporating edge-enabled hubs, IP cameras, voice assistants, motion sensors, and smartphone clients within 60 households recruited across urban and suburban settings. A stratified random sampling approach yields 1200 device-hours of data, with a minimum of 20 households equipped with configurable edge servers and 10 with cloud-centric baselines. Data collection instruments include calibrated benchmark workloads representing multimedia processing, computer vision, and sensor fusion tasks; network trace collectors to capture latency, bandwidth, and jitter; energy meters on representative devices; and a user questionnaire instrument validated for QoE sensitivity and privacy concerns. Additionally, controlled experiments simulate varying network conditions (bandwidth from 5 to 100 Mbps, jitter, and packet loss) to stress-test offloading decisions. Analytical procedures combine quantitative and qualitative techniques. Descriptive statistics summarize baseline performance across offloading configurations. Regression analyses (multivariate linear and panel data) assess the relationship between task features, network conditions, and offloading outcomes (latency, energy, cost). Generalized linear models examine categorical outcomes such as offload decision and reliability. Time-series analysis evaluates performance trends over diurnal cycles. Model-based prediction employs a supervised learning approach, including random forest and gradient boosting, to classify tasks as best-suited for edge, cloud, or local processing, with feature importance metrics to reveal key determinants. ANOVA tests compare performance across heterogeneous households and configurations. The qualitative component utilizes thematic analysis of participant interviews to interpret perceived QoE and privacy trade-offs, with triangulation against quantitative results to enhance validity. The study integrates privacy-preserving evaluation through a threat-model lens, drawing on the Theory of Privacy by Design to examine how architectural choices influence user trust. Expected findings anticipate that edge offloading will reduce latency and energy consumption for compute-intensive, latency-sensitive tasks when network conditions are favorable and privacy requirements are high, whereas cloud offloading may be advantageous for batch-like, high-variance workloads under constrained edge resources. Predictive models are expected to achieve F1-scores above 0.85 in task offloadability classification, with edge-centric policies outperforming static baselines by 20-35% in QoE metrics under fluctuating network conditions. The analysis should reveal a nuanced trade-off between privacy guarantees and performance gains, with user acceptance contingent on transparent privacy controls and predictable service levels. The study contributes to knowledge by providing empirically grounded offloading guidelines, a validated predictive framework for task offload decisions, and a nuanced understanding of how edge infrastructure design shapes user experience in smart homes. The main conclusion posits that adaptive, context-aware offloading policies powered by real-time device and network telemetry substantially improve performance and user satisfaction without compromising privacy, provided that edge resources are provisioning-aware and privacy-by-design principles are embedded. Recommendations include the development of lightweight offloading orchestration layers at the edge, standardized privacy metrics for smart-home data, and deployment guidelines for hybrid edge-cloud architectures that optimize QoE while maintaining security guarantees. The study suggests avenues for further research into reinforcement learning-based offloading, cross-veder interoperability standards, and long-term user-centric evaluations of privacy perception in edge-enabled domestic environments.