Trust-Aware Client Scoring in Federated Learning via Contrastive Attention and Representation Clustering

Federated learning (FL) enables collaborative model training while preserving data privacy, but its decentralized architecture presents significant trust challenges, particularly in the presence of adversarial clients. Most existing defenses rely on prior assumptions, trusted data, or heuristic trust models, thereby limiting their robustness. In this paper, we propose a novel zero-trust scoring mechanism for federated learning that infers client reliability during aggregation, without assuming any prior trust information or access to clean reference data. Our approach assigns trust scores based solely on representation-level information, without access to client data. Each client creates a compact representation via an intra-client self-attention pooling step, and on the server side, an inter-client attention module compares all client representations to produce initial trust scores. These trust scores are then regularized using a hybrid clustering component (K-Means combined with DBSCAN for outlier detection), employing a contrastive cluster-aware loss that promotes consistency within coherent client groups while suppressing the influence of adversarial participants. Experiments on CIFAR-10 and GTSRB demonstrate that the proposed method effectively isolates adversarial behavior, generalizes to unseen clients from the same data distribution, and improves global model performance under diverse poisoningbased adversarial scenarios. This zero-trust, self-organizing framework provides a resilient and practical solution for trustaware aggregation in federated learning.