A Lightweight Trust Framework for Edge AI Systems

Thesis Abstract

In an era increasingly dominated by edge computing, deploying AI workloads on resource-constrained devices raises critical concerns about trust, security, and reliability in autonomous decision-making environments. The study addresses the gap between lightweight, on-device trust mechanisms and the stringent performance constraints of edge AI systems, where traditional centralized trust models incur unacceptable latency and bandwidth costs. The aim is to develop a lightweight trust framework that ensures data integrity, model provenance, and secure collaboration among edge nodes without compromising real-time performance. Specific objectives include (1) to design a modular trust framework tailored for edge AI theorems that integrates lightweight attestation, provenance tracking, and anomaly detection; (2) to formalize a trust metric that combines device capability, data quality, and behavioral evidence into a bounded, interpretable score; (3) to implement a prototype on a heterogeneous edge testbed comprising 50 nodes with Raspberry Pi, NVIDIA Jetson devices, and microcontrollers; (4) to evaluate the framework under realistic workloads including image classification, sensor fusion, and federated learning; and (5) to assess robustness against adversarial interference and network partitioning. The methodology adopts a mixed-methods research design grounded in security and trust theories, notably the Situational Trust Theory and the Formal Provenance Model, augmented by a lightweight Bayesian decision layer. The population consists of edge devices within a smart manufacturing and autonomous robotics setting, with a sample of 50 nodes for validation and a controller node for orchestration. Data collection employs (i) instrumented telemetry logs capturing attestation results, provenance records, and anomaly scores; (ii) synthetic and real-world data streams for vision and sensor fusion tasks; and (iii) a survey of system operators to capture perceived trust levels. Validity and reliability are ensured through triangulation across attestation evidence, provenance integrity checks (log hash chaining and Merkle trees), and cross-validation of anomaly detectors using holdout datasets. Analytical techniques include descriptive statistics to profile trust metrics; regression analysis to quantify the relationship between device characteristics and trust scores; a Bayesian inference layer to update trust in light of new evidence; and ROC/precision-recall metrics to evaluate detector performance. A lightweight Markov decision process (MDP) model specifies policy decisions for resource allocation under varying trust states. The model specification integrates a probabilistic trust propagation mechanism across neighboring edge nodes and uses a constrained optimization to minimize latency while maximizing aggregate trust. Performance evaluation considers throughput, latency, energy consumption, and trust convergence time, with emphasis on scalability to 100–200 nodes in extended scenarios. Ethical considerations address data privacy, secure handling of provenance, and avoidance of unintended bias in trust assessments. Expected findings indicate that the proposed framework achieves sub-100 ms end-to-end attestation latency on edge devices with heterogeneous capabilities, while maintaining a trust score with 95% detection accuracy for compromised or misbehaving nodes. The framework should demonstrate resilience against cascading failures during network partitions, preserving core decision-making capabilities via local provenance and cached trust evidence. The study anticipates that the Bayesian layer will provide robust trust updates with limited data, and the MDP policy will efficiently balance computation, communication, and trust maintenance under resource constraints. The contribution to knowledge comprises (i) a formalized, lightweight trust framework for edge AI that fuses attestation, provenance, and anomaly detection into a cohesive, scalable model; (ii) a demonstrable, reproducible prototype applicable to smart manufacturing, smart cities, and autonomous robotics; and (iii) empirical insights into the trade-offs between trust accuracy, latency, and energy consumption in resource-constrained edge environments. The main conclusion posits that lightweight, modular trust mechanisms can sustain trustworthy edge AI operations without sacrificing performance, provided that trust signals are fused through probabilistic reasoning and policy-aware orchestration. Practical recommendations include guidelines for deploying attestation and provenance in heterogeneous edge ecosystems, calibration of trust thresholds to balance false positives and negatives, and strategies for integrating the framework with federated learning workflows to secure model provenance and collaboration. Suggestions for future work include extending the framework to multi-tenant edge clouds, exploring hardware-assisted trusted execution environments for further minimization of overhead, and empirical validation in large-scale industrial deployments.