DREAM: Disentangled RoI Attention Map for Privacy-Aware Edge Inference
IEEE Transactions on Cognitive Communications and Networking · 2026 · 12 · 9687-9701
DOI: 10.1109/tccn.2026.3706084 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Cognitive Communications and Networking, 2026. Verfasst von Manyou Ma und 3 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Gefördert von: Natural Sciences and Engineering Research Council of Canada, the Digital Research Alliance of Canada (alliancecan.ca), and the National Key R & D Program of Ch….
Abstract (Originalsprache)
In this paper, we consider a communication scenario in which an edge device and an edge server collaboratively perform inference tasks based on an image which is stored at the edge device. The image is considered to be private as it contains sensitive information, making data transmission vulnerable to potential eavesdroppers. For the case when the inference task only requires a specific region-of-interest (RoI) within the image, privacy leakage can be reduced by transmitting only the sub-image within that RoI. We develop a framework for performing RoI segmentation through edge inference, where the segmentation loss and privacy leakage in the transmit feature vector are jointly considered. We propose a deep learning (DL)-based framework called Disentangled REgion-of-interest Attention Map (DREAM), and use the disentanglement approach to ensure that the transmit feature vector does not contain any private information, thereby preventing privacy leakage. We also develop a three-stage adversarial training procedure to guarantee training convergence. Simulations are performed on the CelebAMask-HQ dataset for segmenting the eyes and nose in human faces, while treating gender as the private attribute. The proposed DREAM framework can reduce the accuracy of an eavesdropper’s gender prediction by 38%, while achieving similar segmentation loss compared with two benchmarks.
Förderung
- Natural Sciences and Engineering Research Council of Canada, the Digital Research Alliance of Canada (alliancecan.ca), and the National Key R & D Program of Ch…
Zitieren
APA
Ma, M., Wong, V. W. S., Wang, L., & Schober, R. (2026). DREAM: Disentangled RoI Attention Map for Privacy-Aware Edge Inference. IEEE Transactions on Cognitive Communications and Networking, 12, 9687-9701. https://doi.org/10.1109/tccn.2026.3706084
Vancouver
Ma M, Wong VWS, Wang L, Schober R. DREAM: Disentangled RoI Attention Map for Privacy-Aware Edge Inference. IEEE Transactions on Cognitive Communications and Networking. 2026;12:9687-9701. doi:10.1109/tccn.2026.3706084
BibTeX
@article{Ma2026,
author = {Ma, Manyou and Wong, Vincent W. S. and Wang, Lele and Schober, Robert},
title = {DREAM: Disentangled RoI Attention Map for Privacy-Aware Edge Inference},
journal = {IEEE Transactions on Cognitive Communications and Networking},
year = {2026},
volume = {12},
pages = {9687--9701},
doi = {10.1109/tccn.2026.3706084}
}
EUVIMED