Deepfake face detection based on attention enhancement and multiscale feature fusion

Abstract

Driven by the rapid advancement of generative adversarial networks (GANs), deepfake facial manipulation has emerged as a critical societal concern. Existing detection methods are hindered by rigid multi‑scale feature fusion and insufficient sensitivity to subtle forgery artifacts. To address these challenges, we propose AMF‑Net, an attention‑enhanced multi‑scale fusion network. First, to resolve semantic‑level conflicts and spatial localization biases inherent in traditional feature pyramid networks (FPNs), we design a Conditional Hierarchy Pyramid (CHP‑Neck) that employs a conditional gating fusion mechanism to adaptively calibrate and align multi‑scale features. Second, to boost sensitivity to tampered‑region artifacts, we introduce a multi‑branch attention‑enhanced backbone, MB‑EfficientNetV2, which leverages a parallel multi‑branch architecture with embedded CBAM‑E attention modules to strengthen representational capability for fine‑grained forgery traces. Finally, we construct a lightweight partial self‑attention detection head (LPSA‑Head) that balances global dependency modeling with local detail preservation, thereby improving both detection accuracy and inference efficiency. Experimental results show that our method achieves high accuracy in identifying deepfake faces. Comparative and ablation experiments on widely used benchmarks, including Celeb‑DF (v2), validate its superiority and validity.