Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

IslandRun: Privacy-Aware Multi-Objective Orchestration for Distributed AI Inference

Bala Siva Sai Akhil Malepati

TL;DR

IslandRun addresses the challenge of privacy-aware, multi-objective AI inference across heterogeneous resources by decomposing routing decisions into dedicated agents for privacy, resources, cost, and trust. Its agent-based architecture, coupled with data-locality routing and typed placeholders for cross-boundary contexts, enables per-request decisions that maximize local compute usage while preserving privacy. The framework formalizes a problem of cross-domain, per-request optimization with fail-closed privacy guarantees, and introduces a three-tier island model (personal devices, private edge, public cloud) to balance data sovereignty and scalability. Although primarily theoretical in this work, IslandRun lays a principled foundation for privacy-preserving, decentralized inference orchestration across personal ecosystems and enterprise boundaries, with significant implications for cost, latency, and data governance.

Abstract

Modern AI inference faces an irreducible tension: no single computational resource simultaneously maximizes performance, preserves privacy, minimizes cost, and maintains trust. Existing orchestration frameworks optimize single dimensions (Kubernetes prioritizes latency, federated learning preserves privacy, edge computing reduces network distance), creating solutions that struggle under real-world heterogeneity. We present IslandRun, a multi-objective orchestration system that treats computational resources as autonomous "islands" spanning personal devices, private edge servers, and public cloud. Our key insights: (1) request-level heterogeneity demands policy-constrained multi-objective optimization, (2) data locality enables routing compute to data rather than data to compute, and (3) typed placeholder sanitization preserves context semantics across trust boundaries. IslandRun introduces agent-based routing, tiered island groups with differential trust, and reversible anonymization. This establishes a new paradigm for privacy-aware, decentralized inference orchestration across heterogeneous personal computing ecosystems.

IslandRun: Privacy-Aware Multi-Objective Orchestration for Distributed AI Inference

TL;DR

IslandRun addresses the challenge of privacy-aware, multi-objective AI inference across heterogeneous resources by decomposing routing decisions into dedicated agents for privacy, resources, cost, and trust. Its agent-based architecture, coupled with data-locality routing and typed placeholders for cross-boundary contexts, enables per-request decisions that maximize local compute usage while preserving privacy. The framework formalizes a problem of cross-domain, per-request optimization with fail-closed privacy guarantees, and introduces a three-tier island model (personal devices, private edge, public cloud) to balance data sovereignty and scalability. Although primarily theoretical in this work, IslandRun lays a principled foundation for privacy-preserving, decentralized inference orchestration across personal ecosystems and enterprise boundaries, with significant implications for cost, latency, and data governance.

Abstract

Modern AI inference faces an irreducible tension: no single computational resource simultaneously maximizes performance, preserves privacy, minimizes cost, and maintains trust. Existing orchestration frameworks optimize single dimensions (Kubernetes prioritizes latency, federated learning preserves privacy, edge computing reduces network distance), creating solutions that struggle under real-world heterogeneity. We present IslandRun, a multi-objective orchestration system that treats computational resources as autonomous "islands" spanning personal devices, private edge servers, and public cloud. Our key insights: (1) request-level heterogeneity demands policy-constrained multi-objective optimization, (2) data locality enables routing compute to data rather than data to compute, and (3) typed placeholder sanitization preserves context semantics across trust boundaries. IslandRun introduces agent-based routing, tiered island groups with differential trust, and reversible anonymization. This establishes a new paradigm for privacy-aware, decentralized inference orchestration across heterogeneous personal computing ecosystems.

Paper Structure

This paper contains 61 sections, 6 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. The Multi-Objective Inference Dilemma
  3. Motivation: The Multi-Dimensional Orchestration Challenge
  4. Key Insight: Multi-Objective Agents as First-Class Primitives
  5. Contributions
  6. Paper Organization
  7. Related Work
  8. Inference Serving and Orchestration
  9. Privacy-Preserving Machine Learning
  10. Multi-Objective Optimization
  11. Edge and Hybrid Cloud-Edge Systems
  12. Service Mesh and Traffic Management
  13. Comparative Analysis: IslandRun vs. Existing Systems
  14. Problem Formulation
  15. System Model
  16. ...and 46 more sections

Figures (3)

  • Figure 1: IslandRun architecture. Four agents—WAVES (routing), MIST (privacy), TIDE (resources), LIGHTHOUSE (topology)—synthesize inputs to compute multi-objective scores for each island, then route to SHORE (local) or HORIZON (cloud) based on Pareto dominance. Color encodes roles: blue (routing), purple (privacy), green (resources), teal (coordination), orange/pink (execution targets).
  • Figure 2: Request flow showing route-then-sanitize pipeline. MIST computes sensitivity score, TIDE checks resource availability, and WAVES applies privacy constraint (Pj$\geq$ sr) to filter candidate islands. Personal/trusted islands (P=1.0) bypass MIST sanitization; untrusted routing applies typed placeholders for context preservation.
  • Figure 3: Three-tier island deployment topology. Tier 1 (green): Personal devices with maximum trust (1.0) and zero marginal cost (no MIST required). Tier 2 (yellow): Private edge servers (home servers, VPN endpoints, dedicated edge devices) with moderate trust (0.6-0.8) and fixed operational costs. Tier 3 (red): Public cloud with lower trust (0.3-0.5) requiring MIST sanitization. LIGHTHOUSE maintains mesh connectivity and heartbeat coordination across tiers.

Theorems & Definitions (4)

  • Definition 1: Computing Island
  • Definition 2: Inference Request
  • Definition 3: Privacy Constraint Compliance
  • Definition 4: Trust Boundary Transition