The globalization of the semiconductor supply chain has intensified the proliferation of counterfeit integrated circuits (ICs), posing severe threats to the reliability and security of safety critical systems. Federated Learning (FL) has emerged as a promising privacy preserving paradigm for collaborative counterfeit IC detection, enabling distributed entities to jointly train machine learning models without exchanging sensitive raw data. However, the decentralized nature of FL introduces critical security vulnerabilities. This work first demonstrates that standard FL aggregation defenses are insufficient against a novel Byzantine data poisoning attack. This attack embeds subtle hardware-specific triggers into IC training data, corrupting the global model while evading detection with access to an average of just 4.46% of the dataset. To counteract this vulnerability, we propose Federated Embedding Distribution Authentication (FedEDAuth), a lightweight and proactive authentication framework. FedEDAuth detects and filters malicious clients prior to aggregation by analyzing the statistical properties of client-side feature embeddings against a trusted golden reference. Experimental evaluation across 50 distributed clients and five random seeds demonstrates that FedEDAuth achieves a 100% malicious client detection rate against a stealthy Byzantine data poisoning attack on an IC image dataset. Post-filtering, the federated counterfeit IC detection model achieves an accuracy of 90.94% and an AUC-ROC of 0.91. The results of this experiment underscore the critical necessity of secure client authentication for ensuring resilient collaborative hardware security. While demonstrated on counterfeit IC detection, the architecture is directly applicable to tasks including IoT device attestation, hardware trojan detection, and broader supply-chain integrity verification.