TrustFCL: Task-grained model assessment and aggregation for secure federated continual learning

Federated learning (FL) enables decentralized model training across devices, while federated continual learning (FCL) continuously adapts to evolving tasks (varying in categories, data distributions, or problem domains). Currently, FL security issues, such as model poisoning and privacy leaks, are undermining trust and reliability, prompting the development of defense methods. However, with evolving tasks, when a poisoning attack occurs, traditional FL defense methods only evaluate whether the latest model from each client is poisoned. Therefore, existing approaches face two limitations: (i) defense methods based on static tasks ensure security by restricting parameter mutations, which undermines the model's ability to adapt to new tasks and (ii) knowledge preservation mechanisms (e.g., parameters or gradients), crucial for alleviating forgetting, also introduce new vulnerabilities. Poisoned historical knowledge could be reused to launch attacks that are persistent and difficult to trace. To enable secure and adaptive FCL, this paper proposes TrustFCL, which proposes inter-task reliability and task knowledge reliability as both optimization constraints and control feedback. These reliabilities guide both local anti-forgetting training and global decentralized secure aggregation through a dynamic review committee. Furthermore, TrustFCL stores task reliability information on blockchain to prevent tampering and utilizes the consensus mechanism for privacy-preserving model aggregation. Evaluation results show that TrustFCL reduces the accuracy degradation by 37.1% compared to existing defense methods, and achieves a 17.8% improvement over FCL baselines.