Deepfake detection aims to identify fake videos generated by deep learning techniques. Although deepfake detection methods have achieved remarkable progress, several challenges remain: (1) It is challenging for the existing detection methods to strike a good balance between performance and model complexity. (2) Existing methods often introduce significant redundancy when modeling the relationship between local and global features. (3) Achieving generalization in deepfake detection remains a major challenge, as performance tends to degrade significantly When a disparity exists between the training and testing datasets in terms of their domains. To resolve these issues, we propose an innovative lightweight framework, Tiny Deepfake Detection (TinyDF), which has an extremely small number of parameters and achieves exceptional performance. Specifically, we propose a Pyramid Atrous Aggregation (PAA) module to address the significant redundancy caused by the spatial inconsistency between global and local features. This module enables multiperspective perception of global and local features and maps them into a unified feature space for detailed interaction. Secondly, we propose a low-cost but highly efficient fusion module, Shuffle Fusion Mixer (SFM), which enhances feature interaction through multi-resolution integration, thereby improving cross-dataset generalization. We explore the inaugural application of the Kolmogorov-Arnold idea in the realm of deep forgery detection, which significantly improves the detection performance through powerful nonlinear expression. Extensive experiments demonstrate that TinyDF outperforms existing methods in both generalization and accuracy. Notably, TinyDF requires only 5.38M parameters and 0.59G FLOPs.