AAAI 2026 — Technical Track on Machine Learning

PRISM: Privacy-Aware Routing for Adaptive Cloud–Edge LLM Inference

A context-aware framework that dynamically balances privacy and inference quality through semantic sketch collaboration between cloud LLMs and edge SLMs.

Junfei Zhan1*   Haoxun Shen1*   Zheng Lin2   Tengjiao He3†

1University of Pennsylvania   2University of Hong Kong   3Jinan University
* Equal contribution   † Corresponding author

The Privacy-Performance Dilemma

LLMs excel at understanding natural language, but deploying them raises a fundamental tension between privacy protection and inference quality.

🔒

Privacy Risks in the Cloud

User prompts in sensitive domains (medical, financial) contain personal information. Sending full prompts to cloud LLMs exposes private data to potential breaches.

Edge Device Constraints

Running large models locally preserves privacy but is limited by compute, memory, and energy. Edge SLMs cannot match cloud LLM quality on complex queries.

🔄

Coarse-Grained Protection

Existing approaches apply uniform privacy noise to all tokens regardless of sensitivity, degrading utility unnecessarily for benign content.

How PRISM Works

PRISM executes in four stages, combining entity-level sensitivity profiling with adaptive privacy protection and semantic sketch collaboration.

1

Sensitivity Profiling

The edge device analyzes the user prompt using NER (Named Entity Recognition) to extract entities and compute a category-weighted risk score. Each entity type (PERSON, LOCATION, DATE, etc.) carries a predefined sensitivity weight.

PERSON: 0.9 LOCATION: 0.4 DATE: 0.3 NER + Risk Score
2

Soft Gating Router

An entropy-regularized neural router maps sensitivity indicators to a probability distribution over three execution modes. Unlike hard thresholds, soft gating produces nuanced, context-aware routing decisions.

Edge-Only Collaborative Cloud-Only
3

Adaptive Two-Layer LDP

For collaborative paths, PRISM applies entity-level local differential privacy with two layers: (1) category obfuscation and (2) value perturbation. Privacy budget is adaptively allocated based on each entity's sensitivity weight.

ε-LDP Guarantee Category-Aware Budget Fine-Grained Control
4

Semantic Sketch Collaboration

The cloud LLM receives the perturbed prompt and generates a semantic sketch — a structured outline preserving task structure while hiding sensitive details. The edge SLM then refines this sketch using the original unperturbed context to produce the final response.

Cloud Sketch Generation Edge Refinement Privacy-Preserved Quality

Three Execution Modes

PRISM dynamically routes each query to the optimal execution path based on its privacy risk profile.

☁️

Cloud-Only

Low-risk queries like general knowledge questions are sent directly to the cloud LLM for maximum quality.

PrivacyLow
QualityHigh
🤝

Collaborative

Moderately sensitive queries use adaptive LDP + semantic sketch collaboration for balanced privacy and quality.

PrivacyMedium-High
QualityGood
📱

Edge-Only

Highly sensitive queries (e.g., medical records) are processed entirely on-device for maximum privacy protection.

PrivacyMaximum
QualityModerate

See PRISM in Action

Watch how PRISM processes a privacy-sensitive tourism query through the collaborative pipeline.

User Prompt
"I'm considering traveling to New York in August with my friend Alice. It's a 6-day, 3481 EUR trip. Please help me plan my days off."
🔍
Step 1 — Sensitivity Profiling: Detected 4 entities (PERSON, LOCATION, DATE, MONEY). Risk score: 0.62. Route: Collaborative
🛡️
Step 2 — Adaptive LDP: "New York" → "Bilbury Town", "Alice" → "Emily", "August" → "September", "3481 EUR" → "White trip"
☁️
Step 3 — Cloud Sketch: Cloud LLM generates structured itinerary outline from perturbed prompt (no real user data seen)
Step 4 — Edge Refinement: Edge SLM restores original entities and produces personalized 6-day New York itinerary for Alice

Evaluation Across Four Domains

PRISM is validated on a synthetic, context-rich instruction dataset spanning diverse privacy scenarios.

🏥

Medical

Patient demographics, symptoms, diagnosis history, treatment plans

✈️

Tourism

Travel plans, budgets, destinations, companion details, itineraries

💳

Banking

Transaction histories, card disputes, account information, balances

📚

General Knowledge

Non-sensitive factual queries, benchmarking cloud-only quality

Key Results

PRISM consistently achieves superior privacy-utility trade-offs across all evaluation scenarios.

40-50%
Reduction in energy & latency
vs. baseline LDP methods
ε-LDP
Provable privacy guarantee
with adaptive budget allocation
3
Adaptive execution modes
context-aware routing
4
Evaluation domains
medical, tourism, banking, general

Inference Quality Under Privacy Constraints

PRISM (Ours)Best trade-off
Selective LDPBaseline
Uniform LDPBaseline
Edge-Only (No Privacy Loss)Lower bound

Citation

If you find PRISM useful in your research, please consider citing our paper.

@inproceedings{zhan2026prism,
  title     = {PRISM: Privacy-Aware Routing for Adaptive Cloud-Edge
               LLM Inference via Semantic Sketch Collaboration},
  author    = {Zhan, Junfei and Shen, Haoxun and Lin, Zheng and He, Tengjiao},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {40},
  number    = {33},
  pages     = {28150--28158},
  year      = {2026},
  doi       = {10.1609/aaai.v40i33.40041},
  url       = {https://ojs.aaai.org/index.php/AAAI/article/view/40041}
}