Unified-Copycat: Unifying Heterogeneous Classifiers Using Non Problem Domain Data

Gabriel Braga Ladislau, Jhonatan Machado Leão, Marlon Moratti do Amaral, Claudine Badue, Alberto Ferreira de Souza, Thiago Oliveira-Santos

Abstract: With the success of Convolutional Neural Networks, companies are increasingly offering trained models as paid services through APIs, where users are charged to query different models, without any access to their architecture or training datasets. Additionally, privacy regulations and ethical concerns have also led to the use of black-box models. These constraints make datasets and detailed information about the classifiers entirely unavailable to users. The challenge is further exacerbated when unifying models designed for tasks with different class sets, as class disparities significantly complicate the unification process. Given this scenario, unifying such specialized models into a single classifier provides clear advantages, but existing solutions often assume access to domain data or internal model structures. To address this, we propose a novel approach based on the Copycat framework that unifies heterogeneous models into one unified model capable of inference across all class domains. Unlike prior methods, our approach leverages random natural images from outside the domains of the heterogeneous models, reducing the assumptions required for unification. Comprehensive evaluations show that our method achieves competitive performance with significantly fewer assumptions, marking a step forward in unifying heterogeneous classifiers.

External IDs:dblp:conf/sibgrapi/LadislauLABSO25