<p>Deep face forgery detection mainly aims to address the technology of detecting whether the facial information in images or videos has been forged or changed. The existing deep face forgery detection techniques rely on the visual artifacts and spatial anomalies of the forged images. However, these ways are easily to be affected by factors such as compression ratio, resolution changes, or adversarial perturbations from attackers, which can lead to unstable or significantly reduced detection performance. To solve this issue, this paper creatively proposes a deep lightweight face forgery detection network using multi-scale global features and adaptive weighted channel self-attention. Firstly, by integrating the cross-fusion of global features and multi-scale channel attention technology, a dynamic channel attention mechanism is constructed, which can solve the problem of unstable detection performance. Furthermore, we use depth-wise separable convolution module which was designed to optimize the number of computational parameters, which can meet the compact design and reduce computational requirements effectively. The proposed scheme can achieve an effective balance between high detection accuracy, low computational complexity and parameter-economical, and provides an effective solution for face forgery detection in edge mobile devices and streaming media scenarios. A series of simulation experiments have demonstrated that our method can achieve superior detection performance over multiple classical datasets, and outperforms existing state-of-the-art schemes in terms of detection rate, robustness and complexity.</p>