Askhatuly A., Berdysheva D., Berdyshev A., Adamova A., Yedilkhan D.
CAPG: Context-Aware Perturbation Generation for Multi-Label Adversarial Attacks // Technologies. — 2026. — Vol. 14, No. 4. — Article No. 233. — DOI: 10.3390/technologies14040233.

Abstract: Multi-label deep learning models are widely used in real-world applications where predictions depend on the joint presence of multiple semantically correlated labels. However, existing adversarial attack methods often ignore inter-label dependencies, producing perturbations that are structurally inconsistent or easily detectable. This paper introduces CAPG (Context-Aware Perturbation Generation), a white-box adversarial framework designed to generate selective and contextually consistent perturbations in multi-label settings. The method incorporates correlation-weighted regularization into the adversarial objective, enabling targeted manipulation of specific labels while preserving the contextual integrity of non-target outputs. Experiments on the Pascal VOC 2012 dataset using a ResNet-101 multi-label classifier show that CAPG achieves higher attack success rates and improved contextual consistency compared to FGSM, PGD, CW, and DeepFool under the same perturbation constraints. Additionally, CAPG produces lower perceptual distortion, demonstrating improved structural preservation of adversarial examples and highlighting the importance of correlation-aware robustness evaluation in multi-label deep learning systems.

Link / DOI: https://doi.org/10.3390/technologies14040233