This paper introduces LAA-Net, a novel approach for high-quality deepfake detection. Existing methods often rely on supervised binary classifiers with implicit attention mechanisms, which struggle to generalize to unseen manipulations. To address this, LAA-Net incorporates two main contributions: an explicit attention mechanism within a multi-task learning framework and an Enhanced Feature Pyramid Network (E-FPN). The attention mechanism combines heatmap-based and self-consistency strategies to focus on small, vulnerable regions prone to artifacts. The E-FPN spreads discriminative low-level features into the final output, reducing redundancy. Experiments on multiple benchmarks show that LAA-Net outperforms existing methods in terms of AUC and AP, demonstrating its effectiveness in detecting high-quality deepfakes. The code is available at <https://github.com/10Ring/LAA-Net>.This paper introduces LAA-Net, a novel approach for high-quality deepfake detection. Existing methods often rely on supervised binary classifiers with implicit attention mechanisms, which struggle to generalize to unseen manipulations. To address this, LAA-Net incorporates two main contributions: an explicit attention mechanism within a multi-task learning framework and an Enhanced Feature Pyramid Network (E-FPN). The attention mechanism combines heatmap-based and self-consistency strategies to focus on small, vulnerable regions prone to artifacts. The E-FPN spreads discriminative low-level features into the final output, reducing redundancy. Experiments on multiple benchmarks show that LAA-Net outperforms existing methods in terms of AUC and AP, demonstrating its effectiveness in detecting high-quality deepfakes. The code is available at <https://github.com/10Ring/LAA-Net>.