The increasing realism of Deepfake media generated by modern models poses significant risks to digital trust and forensics. Although Vision Transformer-based detectors outperform Convolutional Methods, they remain vulnerable to unstable attention under common perturbations and often operate in closed-set settings, leading to overconfidence on unseen forgeries. This work analyzes attention consistency as an information-theoretic approach to enhance detection reliability. We propose an Attention-Consistent Vision Transformer (ViT-ACR) that employs Kullback–Leibler divergence to stabilize self-attention distributions. An entropy-based decision rule enables uncertainty-aware open-set inference. Experiments under identity-disjoint protocols across multiple public benchmarks demonstrate strong cross-dataset generalization, achieving 99.28% accuracy on FaceForensics++ and 89.57% mean AUC on six unseen datasets, with reduced predictive uncertainty under compression.