.png){kind=icon}
.png){kind=icon}
Return
Continual face forgery detection based on relation-aware spatial-frequency interaction aggregation and contrastive learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.112456.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the proliferation of face forgery images on the internet, there is an increasing interest in designing effective methods for fake image detection. Most existing methods exhibit poor generalization performance, which makes it difficult to adapt to emerging forgery techniques. To address this challenge, we propose a novel continual face forgery detection (CFFD) framework that integrates multi-view knowledge distillation and a hybrid sampling replay mechanism to improve the generalization of the model for evolving forgery techniques. Within the framework, we also present a relation-aware spatial-frequency interaction aggregation network (RSIA-Net). This network utilizes the designed relation-aware spatial-frequency interaction aggregation (RSIA) modules to adaptively reweight the spatial and frequency domain enhancement features based on relevance-guided information and perform interaction aggregation. This hierarchical relevance-guided refinement mechanism helps the model extract more fine-grained representations. Furthermore, we propose a hierarchical spatial-frequency contrastive learning mechanism (HSCL), facilitating the dual-domain information fusion and learning an embedding space with enhanced intra-class consistency and inter-class diversity by modeling intra-domain and cross-domain feature correlations on multi-level features in the spatial and frequency domains. Extensive experiments on four public datasets demonstrate that the proposed method has superior performance compared to the most advanced techniques in detecting different types of face forgeries.
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
View more organizations