hh.sePublications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Efficient Obfuscation Techniques for Privacy-Preserving Machine Learning in IoT Edge Devices: Balancing Privacy and Utility through Adaptive Obfuscation in Edge-Based Machine Learning
Halmstad University, School of Information Technology.
Halmstad University, School of Information Technology.
2025 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
Abstract [en]

The rapid growth of Internet of Things (IoT) ecosystems in healthcare, autonomous systems, and industrial automation has enabled powerful capabilities but also heightened privacy risks. Edge devices often capture visual data containing sensitive attributes such as identity, background, or biometric traits. Traditional privacy-preserving methods, such as homomorphic encryption or differential privacy, offer strong guarantees but are too computationally demanding for real–time use on resource-constrained devices.

This thesis proposes an inference–time obfuscation framework tailored for IoT edge deployment. The approach employs an adversarial encoder–decoder with Gradient Reversal Layers (GRL) and supervised contrastive learning to disentangle utility–relevant from privacy–sensitive features, enabling selective suppression of attributes such as color, digit shape, or background while preserving task–critical utility. The framework is evaluated on Colored-MNIST and Cityscapes, with privacy leakage assessed through dedicated attack models (ColorNet, RecNet). Results demonstrate that the method significantly reduces leakage of private attributes while maintaining high utility accuracy, even under constrained bottleneck sizes, making it practical for real-time IoT edge applications.

Abstract [sv]

Den snabba tillväxten av Internet of Things (IoT) inom hälso- och sjukvård, autonoma system och industriell automation har möjliggjort kraftfulla funktioner men också ökat integritetsriskerna. Edge enheter fångar ofta visuella data som innehåller känsliga attribut, såsom identitet, bakgrund eller biometriska kännetecken. Traditionella metoder för integritetsskydd, såsom homomorf kryptering eller differential privacy, erbjuder starka garantier men är alltför beräkningskrävande för realtidsanvändning på resursbegränsade enheter.

Denna avhandling föreslår ett obfuskationsramverk för inferenstid, särskilt anpassat för IoT-edge-miljöer. Metoden bygger på en adversariell encoder–decoder med Gradient Reversal Layers (GRL) och övervakad kontrastiv inlärning för att särskilja nyttorelevanta från integritetskänsliga egenskaper. Detta möjliggör selektiv undertryckning av attribut såsom färg, sifferform eller bakgrund, samtidigt som uppgiftskritisk nytta (t.ex. objekt- eller sifferidentitet) bevaras. Ramverket utvärderas på Colored-MNIST och Cityscapes, där integritetsläckage analyseras med hjälp av dedikerade attackmodeller (ColorNet, RecNet). Resultaten visar att metoden påtagligt minskar läckage av privata attribut samtidigt som hög nyttokvalitet bibehålls, även vid begränsade bottleneck-storlekar, vilket gör den praktiskt användbar för realtidsapplikationer på IoT-edge-enheter.

Place, publisher, year, edition, pages
2025. , p. 85
Series
Forskning i Halmstad, ISSN 1400-5409
Keywords [en]
Privacy-Preserving Machine Learning, Image Obfuscation, Gradient Reversal Layer (GRL), Contrastive Learning, Utility-Preserving Data Transformation (UPDT), IoT Edge Devices, Representation Learning, Adversarial Training, Semantic Segmentation, ColorNet / RecNet Evaluation
National Category
Computer Systems Computer Engineering Computer Vision and Learning Systems
Identifiers
URN: urn:nbn:se:hh:diva-57418OAI: oai:DiVA.org:hh-57418DiVA, id: diva2:2001342
Subject / course
Computer science and engineering
Educational program
Computer Science and Engineering, 300 credits
Presentation
2025-08-19, E208, Kristian IV:s väg 3, 301 18, Halmstad, 09:00 (English)
Supervisors
Examiners
Available from: 2025-09-29 Created: 2025-09-26 Last updated: 2025-10-01Bibliographically approved

Open Access in DiVA

fulltext(6438 kB)220 downloads
File information
File name FULLTEXT02.pdfFile size 6438 kBChecksum SHA-512
6cf05b47f8b1c31f2c3f70636a09a97f8d7d52efae65222052b4928e84e758174105136806dbc0c736126c9ea899304c578d08188fcde0c66ddc52f3b915d6f4
Type fulltextMimetype application/pdf

By organisation
School of Information Technology
Computer SystemsComputer EngineeringComputer Vision and Learning Systems

Search outside of DiVA

GoogleGoogle Scholar
Total: 221 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 2132 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf