This paper proposes a method called Patch-wise Adversarial Regularization (PAR) to train convolutional neural networks that are robust to domain shifts. The method penalizes the predictive power of local representations in earlier layers, forcing the network to rely on global structure rather than superficial patterns like color and texture. The approach introduces a patch-wise classifier that predicts labels based on local features, while the main network is optimized to fool these classifiers while maintaining accurate predictions at the final layer. The method is evaluated on various synthetic and real-world datasets, including ImageNet-Sketch, a new benchmark consisting of sketch-like images. Results show that PAR improves generalization across different domains, especially when domain information is not available. The method is implemented efficiently using 1×1 convolutions and is tested on multiple variants, including more powerful classifiers and broader local patterns. Experiments demonstrate that PAR outperforms existing methods in domain adaptation and generalization tasks, particularly in scenarios where the model must generalize to new, unseen domains. The paper also highlights the importance of learning global concepts over local patterns for robust visual recognition.This paper proposes a method called Patch-wise Adversarial Regularization (PAR) to train convolutional neural networks that are robust to domain shifts. The method penalizes the predictive power of local representations in earlier layers, forcing the network to rely on global structure rather than superficial patterns like color and texture. The approach introduces a patch-wise classifier that predicts labels based on local features, while the main network is optimized to fool these classifiers while maintaining accurate predictions at the final layer. The method is evaluated on various synthetic and real-world datasets, including ImageNet-Sketch, a new benchmark consisting of sketch-like images. Results show that PAR improves generalization across different domains, especially when domain information is not available. The method is implemented efficiently using 1×1 convolutions and is tested on multiple variants, including more powerful classifiers and broader local patterns. Experiments demonstrate that PAR outperforms existing methods in domain adaptation and generalization tasks, particularly in scenarios where the model must generalize to new, unseen domains. The paper also highlights the importance of learning global concepts over local patterns for robust visual recognition.