Opportunistic Browser-Edge AI: Dynamic Client-Cloud Offloading under Device Constraints and Network Heterogeneity

The emergence of WebAssembly (WASM) and WebGPU runtimes has enabled complex deep learning models to execute directly within consumer web browsers without client-side installation. However, browser-based edge inference suffers from unique sandbox constraints: constrained single-precision memory limits, main-thread rendering stalls during tensor computation, and energy penalties on battery-powered mobile devices. Conversely, offloading inference entirely to cloud servers introduces wide-area network (WAN) round-trip delays, bandwidth contention, and severe data privacy vulnerabilities. In this paper, we present the Adaptive Web-Edge AI Framework, an opportunistic client-cloud inference runtime for browser environments. Our framework introduces three core contributions: (1) an isolated, off-main-thread Dedicated Web Worker architecture that decouples deep learning tensor kernels from DOM rendering, guaranteeing 60/120 FPS UI responsiveness; (2) a Passive EWMA Telemetry Engine with an automated circuit breaker that eliminates the 42–88 ms active probe delay inherent in legacy offloading systems; and (3) an adaptive, multi-objective constrained routing policy that dynamically schedules workloads across local WebGPU/WASM engines and remote cloud servers based on sequence length, battery state, network profiles, and strict privacy guarantees. We evaluate our system across 3,560 empirical trials spanning three distinct AI modalities (DistilBERT sentiment analysis, MiniLM semantic vector embeddings, and DistilGPT-2 text generation) and six standardized wireless and wireline network profiles (Fiber, 5G, Wi-Fi, 4G, 3G, and Geostationary Satellite). Our empirical results prove: (i) on short workloads, local WebGPU inference is statistically significantly faster than cloud execution (p = 3.65 × 10⁻⁴, Cohen’s d = 0.94), while on long workloads (L > 800 characters) cloud offloading achieves a 2.2× speedup (p = 1.51 × 10⁻¹¹, Cohen’s d = 12.31); (ii) for dense semantic vector embeddings, local edge execution outperforms remote cloud transfers by 18.1× across cellular networks; (iii) passive EWMA telemetry cuts turnaround latency by 65.8% (p = 3.14 × 10⁻¹²⁹); and (iv) strict privacy constraints are enforced with 0.0% data leakage under edge failure modes. Index Terms—Edge AI, WebAssembly, WebGPU, Computation Offloading, Client-Cloud Systems, Telemetry, Privacy-Preserving AI.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24320698
Primary Topic
IoT and Edge/Fog Computing
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Opportunistic Browser-Edge AI: Dynamic Client-Cloud Offloading under Device Constraints and Network Heterogeneity

Vibhav J Bharadwaj
2 links
Zenodo (CERN European Organization for Nuclear Research)
IoT and Edge/Fog Computing
preprint

Opportunistic Browser-Edge AI: Dynamic Client-Cloud Offloading under Device Constraints and Network Heterogeneity

Vibhav J Bharadwaj
preprint en

Abstract

The emergence of WebAssembly (WASM) and WebGPU runtimes has enabled complex deep learning models to execute directly within consumer web browsers without client-side installation. However, browser-based edge inference suffers from unique sandbox constraints: constrained single-precision memory limits, main-thread rendering stalls during tensor computation, and energy penalties on battery-powered mobile devices. Conversely, offloading inference entirely to cloud servers introduces wide-area network (WAN) round-trip delays, bandwidth contention, and severe data privacy vulnerabilities. In this paper, we present the Adaptive Web-Edge AI Framework, an opportunistic client-cloud inference runtime for browser environments. Our framework introduces three core contributions: (1) an isolated, off-main-thread Dedicated Web Worker architecture that decouples deep learning tensor kernels from DOM rendering, guaranteeing 60/120 FPS UI responsiveness; (2) a Passive EWMA Telemetry Engine with an automated circuit breaker that eliminates the 42–88 ms active probe delay inherent in legacy offloading systems; and (3) an adaptive, multi-objective constrained routing policy that dynamically schedules workloads across local WebGPU/WASM engines and remote cloud servers based on sequence length, battery state, network profiles, and strict privacy guarantees. We evaluate our system across 3,560 empirical trials spanning three distinct AI modalities (DistilBERT sentiment analysis, MiniLM semantic vector embeddings, and DistilGPT-2 text generation) and six standardized wireless and wireline network profiles (Fiber, 5G, Wi-Fi, 4G, 3G, and Geostationary Satellite). Our empirical results prove: (i) on short workloads, local WebGPU inference is statistically significantly faster than cloud execution (p = 3.65 × 10⁻⁴, Cohen’s d = 0.94), while on long workloads (L > 800 characters) cloud offloading achieves a 2.2× speedup (p = 1.51 × 10⁻¹¹, Cohen’s d = 12.31); (ii) for dense semantic vector embeddings, local edge execution outperforms remote cloud transfers by 18.1× across cellular networks; (iii) passive EWMA telemetry cuts turnaround latency by 65.8% (p = 3.14 × 10⁻¹²⁹); and (iv) strict privacy constraints are enforced with 0.0% data leakage under edge failure modes. Index Terms—Edge AI, WebAssembly, WebGPU, Computation Offloading, Client-Cloud Systems, Telemetry, Privacy-Preserving AI.

Zenodo (CERN European Organization for Nuclear Research)
IoT and Edge/Fog Computing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon