EUVIMEDEUVIMED entdecken
Crossref

RobuGCL: Stability-Aware Graph Contrastive Learning Against AI-Enabled Attacks in Consumer Electronics

Adnan Ali, Jinlong Li, Farooq Ahmed, Hai Xie, Rayan Hamza Alsisi

IEEE Transactions on Consumer Electronics · 2026-02 · 72 (1) · 1532-1543

DOI: 10.1109/tce.2025.3649554 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Consumer Electronics, 2026-02. Verfasst von Adnan Ali und 4 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Basic and Applied Basic Research Foundation of Guangdong Province, National Natural Science Foundation of Hainan Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Graph contrastive learning (GCL) is a powerful paradigm for self-supervised representation learning on consumer electronics (CE) graphs. Conventional GCL lacks temporal, multi-channel adversary modeling and typically relies on uniform, all-node negatives, making it fragile under CE adversarial conditions and memory- and latency-intensive on-device budgets. We propose RobuGCL , a stability-aware framework designed for CE memory, compute, and latency constraints. RobuGCL integrates two co-designed components: (i) a Temporal Adversarial Contrastive Attack (TACA), which periodically applies worst-case structure, feature, and projection perturbations together with a stability regularizer to align clean and perturbed embeddings; and (ii) a budgeted, adaptive negative sampler , which partitions negatives into easy/medium/hard pools and expands the sampled set only when the training loss plateaus, under a strict 50% global cap, effectively halving the worst-case contrastive denominator versus all-negatives. This design allows the contrastive objective to respond dynamically to attack-induced distribution shifts while bounding peak memory compared to all-negatives baselines. We evaluate RobuGCL on nine public benchmarks: the clean-training variant (RobuGCL c ) on all nine and the full adversarial model on eight (DBLP is excluded during the adversarial pass for GPU memory). Relative to the strongest published baseline per dataset, the full RobuGCL achieves state-of-the-art micro-F1 on 7/8 adversarially trained datasets (and 8/9 when including RobuGCL c on DBLP), with gains up to + 2.40 percentage points. Under unseen attacks at 35% budget (RANDOM, DICE, DICE d ), RobuGCL incurs $\le $ 1.0% relative F1-Drop on five of six datasets for deletion-only DICE d , while mixed DICE (additions + deletions) is more disruptive. By jointly addressing robustness and efficiency, RobuGCL advances practical graph representation learning for CE applications such as recommendation, personalization, and IoT analytics.

Förderung

  • National Natural Science Foundation of China
  • Basic and Applied Basic Research Foundation of Guangdong Province
  • National Natural Science Foundation of Hainan Province

Zitieren

APA

Ali, A., Li, J., Ahmed, F., Xie, H., & Alsisi, R. H. (2026). RobuGCL: Stability-Aware Graph Contrastive Learning Against AI-Enabled Attacks in Consumer Electronics. IEEE Transactions on Consumer Electronics, 72(1), 1532-1543. https://doi.org/10.1109/tce.2025.3649554

Vancouver

Ali A, Li J, Ahmed F, Xie H, Alsisi RH. RobuGCL: Stability-Aware Graph Contrastive Learning Against AI-Enabled Attacks in Consumer Electronics. IEEE Transactions on Consumer Electronics. 2026;72(1):1532-1543. doi:10.1109/tce.2025.3649554

BibTeX
@article{Ali2026,
  author = {Ali, Adnan and Li, Jinlong and Ahmed, Farooq and Xie, Hai and Alsisi, Rayan Hamza},
  title = {RobuGCL: Stability-Aware Graph Contrastive Learning Against AI-Enabled Attacks in Consumer Electronics},
  journal = {IEEE Transactions on Consumer Electronics},
  year = {2026},
  volume = {72},
  number = {1},
  pages = {1532--1543},
  doi = {10.1109/tce.2025.3649554}
}
Kostenlos starten →