An Adaptive Multi-Stage Computational Framework for Secure Event Recognition in Resource-Constrained Distributed Environments
Main article
Abstract
The proliferation of sensing devices at the network edge has created an urgent need for event-recognition systems that operate accurately and securely under severe constraints on energy, memory, and communication bandwidth. Conventional approaches force an unsatisfying choice: cloud-centric pipelines achieve high accuracy but incur prohibitive latency, energy, and privacy costs, whereas purely on-device inference sacrifices accuracy on the difficult, safety-critical events that matter most. This paper presents the Adaptive Multi-Stage Framework (AMSF), a computational architecture that distributes recognition across a three-tier edge-fog-cloud hierarchy and routes each sensor observation only as far up the hierarchy as its difficulty requires. A lightweight on-device classifier resolves the large majority of routine observations locally; a confidence-gated escalation mechanism forwards only ambiguous cases to a compact fog-tier model and, in rare instances, to a cloud-tier ensemble. We formalize the per-sample routing problem as a constrained optimization that balances expected accuracy against expected energy and latency, and we derive adaptive confidence thresholds that adjust to workload and channel conditions at run time. To protect the privacy of continuously sensed data, model updates are exchanged through a secure aggregation protocol combined with calibrated differential-privacy noise and pairwise masking, so that the coordinating server never observes any individual update in the clear. We evaluate AMSF on three public human-activity and anomaly-recognition benchmarks under realistic device profiles. AMSF attains 94.3% accuracy while consuming only 6.3 mJ per inference, recovering 97.8% of the accuracy of a cloud-only oracle at roughly one-sixth of its energy cost and one-half of its end-to-end latency, and it degrades gracefully as the number of participating nodes grows from ten to four hundred. An ablation study isolates the contribution of each component, and a privacy-utility analysis shows that strong formal privacy is attainable with less than one percentage point of accuracy loss. The results indicate that difficulty-aware, security-first computation offers a practical path toward trustworthy distributed event recognition.
