An adaptive spatio-temporal framework for carbon-aware and privacy-preserving DNN partitioning

Introduction Choosing where to split a deep neural network (DNN) between device and cloud requires balancing latency, carbon emissions, privacy and device thermal state, which existing partitioning methods rarely address jointly. Methods We propose GreenSplit, an adaptive spatio-temporal framework that selects the split layer and cloud region by minimising a normalised four-term cost under a hard privacy exclusion zone, with an Adaptive Weight Controller (AWC), current and forecast carbon intensity, Laplace noise on transmitted activations, a thermal-battery model, and a dynamic voltage and frequency scaling (DVFS) safeguard. We evaluated it in simulation on MobileNetV2 across three cloud regions (200 sustained-load trials), extended the decision model to ResNet-50, ViT-Base and the CLIP ViT-B/32 image encoder, and tested the noise mechanism on a pretrained ImageNet MobileNetV2 with 1,000 real images. Results Under our assumed energy constants, offloading a lightweight CNN is carbon-costly (5.8–6.7 × on-device carbon), so the framework keeps MobileNetV2 local; it then behaves as local execution with DVFS, emitting 66.0% of the carbon of unmanaged local execution with no privacy-zone violations (4% for baselines without the zone). Its 32.5% throttle rate at nominal thermal parameters is not robust (0–100% over plausible parameters), although DVFS lowers throttling in every draw. Laplace noise at operating budgets (ϵ r ≤ 10 leaves the cloud classifying at chance, while accuracy-preserving budgets ( ϵ r ≈ 3 × 1 0 4 –3 × 10 5 ) allow perfect re-identification among 1,000 images and give a vacuous differential-privacy bound. The AWC's benefit in its original scenario does not transfer to unseen disturbances in its original min–max form; a fixed-scale variant is stable but gives only a modest gain and no thermal benefit. Discussion The hard exclusion zone, not the noise, provides utility-preserving protection. All results are simulated; validation on physical devices is required.

Authors

Institutions

Publication Details

Journal
Frontiers in Artificial Intelligence
Published
2026-11-17
DOI
https://doi.org/10.3389/frai.2026.1961682
Primary Topic
Green IT and Sustainability
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An adaptive spatio-temporal framework for carbon-aware and privacy-preserving DNN partitioning

Syamasudha Veeragandham, SAANVI GOEL
2 links
Frontiers in Artificial Intelligence
Green IT and Sustainability
article

An adaptive spatio-temporal framework for carbon-aware and privacy-preserving DNN partitioning

Syamasudha Veeragandham, SAANVI GOEL
article en

Abstract

Introduction Choosing where to split a deep neural network (DNN) between device and cloud requires balancing latency, carbon emissions, privacy and device thermal state, which existing partitioning methods rarely address jointly. Methods We propose GreenSplit, an adaptive spatio-temporal framework that selects the split layer and cloud region by minimising a normalised four-term cost under a hard privacy exclusion zone, with an Adaptive Weight Controller (AWC), current and forecast carbon intensity, Laplace noise on transmitted activations, a thermal-battery model, and a dynamic voltage and frequency scaling (DVFS) safeguard. We evaluated it in simulation on MobileNetV2 across three cloud regions (200 sustained-load trials), extended the decision model to ResNet-50, ViT-Base and the CLIP ViT-B/32 image encoder, and tested the noise mechanism on a pretrained ImageNet MobileNetV2 with 1,000 real images. Results Under our assumed energy constants, offloading a lightweight CNN is carbon-costly (5.8–6.7 × on-device carbon), so the framework keeps MobileNetV2 local; it then behaves as local execution with DVFS, emitting 66.0% of the carbon of unmanaged local execution with no privacy-zone violations (4% for baselines without the zone). Its 32.5% throttle rate at nominal thermal parameters is not robust (0–100% over plausible parameters), although DVFS lowers throttling in every draw. Laplace noise at operating budgets (ϵ r ≤ 10 leaves the cloud classifying at chance, while accuracy-preserving budgets ( ϵ r ≈ 3 × 1 0 4 –3 × 10 5 ) allow perfect re-identification among 1,000 images and give a vacuous differential-privacy bound. The AWC's benefit in its original scenario does not transfer to unseen disturbances in its original min–max form; a fixed-scale variant is stable but gives only a modest gain and no thermal benefit. Discussion The hard exclusion zone, not the noise, provides utility-preserving protection. All results are simulated; validation on physical devices is required.

Frontiers in Artificial IntelligenceVol. 9
Vellore Institute of Technology University (IN)
Climate action
Green IT and Sustainability
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