FedACK: Federated Adversarial Contrastive Knowledge Distillation for Cross-Lingual and Cross-Model Social Bot Detection

Abstract

Social bot detection is of paramount importance to the resilience and\nsecurity of online social platforms. The state-of-the-art detection models are\nsiloed and have largely overlooked a variety of data characteristics from\nmultiple cross-lingual platforms. Meanwhile, the heterogeneity of data\ndistribution and model architecture makes it intricate to devise an efficient\ncross-platform and cross-model detection framework. In this paper, we propose\nFedACK, a new federated adversarial contrastive knowledge distillation\nframework for social bot detection. We devise a GAN-based federated knowledge\ndistillation mechanism for efficiently transferring knowledge of data\ndistribution among clients. In particular, a global generator is used to\nextract the knowledge of global data distribution and distill it into each\nclient's local model. We leverage local discriminator to enable customized\nmodel design and use local generator for data enhancement with hard-to-decide\nsamples. Local training is conducted as multi-stage adversarial and contrastive\nlearning to enable consistent feature spaces among clients and to constrain the\noptimization direction of local models, reducing the divergences between local\nand global models. Experiments demonstrate that FedACK outperforms the\nstate-of-the-art approaches in terms of accuracy, communication efficiency, and\nfeature space consistency.\n