Text this: In-the-wild deepfake detection using adaptable CNN models with visual class activation mapping for improved accuracy