CLTBA: learnable trigger backdoor attack via frequency domain injection

Deep neural networks have achieved remarkable success in image classification, but their training processes remain vulnerable to backdoor attacks. Existing frequency-domain backdoor attacks typically rely on manually designed and data-independent triggers, limiting their adaptability across different models and datasets. Moreover, fixed spectral perturbations may introduce detectable frequency anomalies. To address these limitations, we propose CLTBA, a backdoor attack based on learnable frequency-domain triggers. CLTBA jointly optimizes the trigger distribution and model parameters through an adaptive spectral mask and a dynamic training strategy, enabling flexible trigger generation while preserving visual quality. A spectral preservation mechanism further constrains trigger perturbations within a controllable frequency region, thereby reducing spectral anomalies and improving stealthiness. Experiments on GTSRB, CelebA, and MNIST with ResNet-18 and EfficientNet-B0 demonstrate that CLTBA achieves high attack success rates while maintaining benign accuracy and shows improved resistance to representative backdoor defenses and analysis methods, including STRIP, Neural Cleanse, Fine-Pruning, and Grad-CAM.