Pixel-Depth Prototypical Knowledge Consolidation Network for Deepfake Detection

The rapid development of deepfake technology poses a serious threat to digital trust, making reliable detection increasingly important. Most existing methods rely on a uni-model isolated extracted feature for detection, which limits their ability to capture the complementary geometric cues. Moreover, many detectors are trained on fixed datasets and lack mechanisms to retain or reuse learned knowledge, reducing their adaptability to new or evolving forgery types. In this paper, we propose Pixel-Depth Prototypical Knowledge Consolidation Network (PDPKC-Net), which jointly utilizes pixel information and depth-based geometric cues to learn richer forgery-related representations. Specifically, we design a dual-branch Pixel-Depth Synergistic Encoder (PDSE) to extract complementary appearance and geometric features, and introduce a Prototypical Knowledge Memory (PKM) to store, refine, and query shared forgery patterns for enhanced detection stability. Experimental results on several deepfake benchmarks show that the proposed framework achieves strong and consistent performance, demonstrating that combining multi-modal feature learning with knowledge consolidation is an effective strategy for improving deepfake detection.