Abstract
The unique artistic style is crucial to artists' occupationalcompetitiveness, yet prevailing Art Commission Platforms rarely supportstyle-based retrieval. Meanwhile, the fast-growing generative AI techniquesaggravate artists' concerns about releasing personal artworks to publicplatforms. To achieve artistic style-based retrieval without exposing personalartworks, we propose FedStyle, a style-based federated learning crowdsourcingframework. It allows artists to train local style models and share modelparameters rather than artworks for collaboration. However, most artistspossess a unique artistic style, resulting in severe model drift among them.FedStyle addresses such extreme data heterogeneity by having artists learntheir abstract style representations and align with the server, rather thanmerely aggregating model parameters lacking semantics. Besides, we introducecontrastive learning to meticulously construct the style representation space,pulling artworks with similar styles closer and keeping different ones apart inthe embedding space. Extensive experiments on the proposed datasets demonstratethe superiority of FedStyle.