Semantic Subgraph Extraction Attacks To Convolutional Neural Networks

Existing adversarial machine learning (AML) methods typically require substantial computational resources. In this work, we investigate how class-level semantic relationships can be exploited to influence adversarial attack performance. Specifically, we first aim to quantify the effect of known factors, such as semantic similarity–defined as the degree of shared intrinsic attributes–on attack efficiency. Experiments on CIFAR-100 and Tiny-ImageNet using VGG and ResNet architectures show that targeting semantically similar classes can reduce perturbation magnitudes and iterations by up to 41% and 23%, respectively. Motivated by these findings, we introduce a lightweight, one-shot, semantic subgraph extraction attack (SSEA), which constructs semantic subgraphs by leveraging class-level semantic relationships between source and adversarial target classes. Our method extracts subgraphs in a single inference pass, requires no fine-tuning or external models, preserves the original network weights, and integrates seamlessly into any white-box attack scenario. On CIFAR-100, SSEA improves the Top-1 attack success rate (ASR) by up to 8.17% for PGD on VGG-19 and nearly doubles the effectiveness of the Jitter attack. Additionally, our approach reduces floating-point operations and model size by up to 18% and 42%, respectively.