Collusion-Based Poisoning Attacks Against Blockchained Federated Learning

The emerging metaverses can be regarded as the current wave of the digital civilization revolution due to it build a new connection between the physical and virtual space, and such connections make it possible for virtual space to interact with real word. However, some concerns, such as privacy data leakage incidents, still hinder the widespread deployment of the metaverse. The blockchained federated learning (FL) architecture can train a federated global model towards multiple stakeholders without the need for centralized management servers and the risk of privacy data leaking as a new paradigm to address the metaverse dilemma. Although many studies have been proposed for the great development potential of blockchain and federated learning for metaverse, few studies, if not none, have explored security threat against the blockchained FL architecture. This work report a prominent phenomenon that the combination of FL and blockchain may pose a more serious security threat, and first to propose Collusion-based Poisoning Attacks (CPA) in the blockchained FL architecture. To be specific, the malicious participants manipulate the model training process so that poisoned global models are aggregated with the most significant probability during the FL training phase. Then submits to the blockchain system and enables the poisoned models to be affected for all participants due to the strong consistency of blockchain. Experimental analysis demonstrates that the proposed CPA can achieve superior performance on multiple aspects and has better performances than the state-of-the-art attack method on stealth and robustness. This research can shed light on the hidden vulnerabilities of the blockchained FL architecture and lay the foundation for build a more open and secure metaverse.