PHA-NAS: Perceptual Hashing for Video Content Authentication Based on Neural Architecture Search

With the rapid growth of edited video content, perceptual video hashing has become increasingly important for content authentication. Most existing methods often rely on predefined feature extraction techniques, which limits their ability to comprehensively characterize complex video content. To address this issue, this paper proposes PHA-NAS, a perceptual video hashing method based on neural architecture search (NAS) that automatically discovers effective architecture for video authentication task. It adopts a dual branch structure, where the fast branch focuses on temporal dynamics and the slow branch captures spatial features. A search space is then constructed for both branches with candidate operations of different receptive fields, allowing it to adaptively select suitable receptive fields and enhance multi-scale feature representation. After the search process identifies the optimal operation at each stage, the resulting architecture is reconstructed to form the proposed PHA-NAS. Furthermore, a chained joint cost estimation algorithm is designed to approximate the search cost while jointly constraining both hash generation and computational complexity. Experimental results demonstrate that PHA-NAS achieves improved video authentication accuracy compared with state-of-the-art methods, yielding an increase of 0.0048 in the area under the ROC curve, while maintaining a favorable balance between robustness and discrimination capability.