Adaptive and robust personalized federated learning with knowledge-driven logits interaction for secure edge intelligence

As edge intelligence increasingly requires collaborative training, data-driven decision-making, and privacy protection, personalized federated learning (PFL) has become pivotal in addressing data silos and safeguarding sensitive information. Yet, existing models struggle with generalization, privacy adaptability across heterogeneous clients, and communication efficiency. To overcome these challenges, a Personalized Trusted Federated Learning (PT-FL) algorithm is proposed utilizing a logit interaction architecture and knowledge cache. PT-FL minimizes communication overhead by avoiding direct parameter exchange between devices, thereby enhancing privacy on edge devices. The server-side knowledge cache captures recent insights from each client’s private data, guiding client model updates and effectively mitigating Byzantine attacks and collusion. We have formally verified a set of invariants that assures correctness of asynchronous and concurrent cache operations, enabling safe, scalable, and composable deployment of our neighbor-based inference caching systems. Additionally, knowledge distillation via the cache supports personalized client model training. An extensive experiment demonstrate PT-FL’s superior performance. Under normal conditions, PT-FL achieves high accuracy with convergence 1.8 times faster than comparative algorithms. Moreover, under adversarial conditions, PT-FL maintains accuracy at 72.8%, surpassing baseline methods by 19.7% to 74.6%, thus validating its robustness against attacks through adaptive contamination suppression.