This work investigates a multi-attribute deepfake detection architecture that combines attention-guided feature generation with additional supervisory signals from camera sensor fingerprints. The approach is designed to improve generalization across unseen datasets and manipulation attacks.