Connected Autonomous Vehicles (CAVs) rely on knowledge exchange to improve situational awareness and coordination. However, centralized or insecure sharing of semantic data exposes critical privacy and trust risks. We present a decentralized framework that enables privacy-preserving knowledge graph sharing and robust federated learning among CAVs, without relying on trusted coordinators. Each vehicle constructs a local semantic graph from environmental observations, privatized through randomized encoding and ephemeral encryption to preserve sensitive relationships. Model updates are secured by Binius-based Zero-Knowledge Proofs (ZKPs), providing lightweight, non-interactive cryptographic verification of update integrity. A lightweight blockchain anchors proof commitments for tamper-resistance, while decentralized reputation scores adaptively filter participants based on verifiable trustworthiness. Empirical evaluations with \(N=100\) nodes, including \(20\%\) malicious actors, show that our framework reduces poisoning attack success rates from \(69.94\%\) (standard FL) to \(65.15\%\) , with less than \(1.5\%\) degradation in final model accuracy. Proof generation and verification incur only 0.123 s per update, and communication overhead grows modestly from 35 KB to 45 KB per round—remaining fully compatible with 5G vehicular networks. Knowledge graph membership inference attacks succeed with probability below \(0.5\%\) even under auxiliary knowledge assumptions. An ablation study confirms that resilience arises from the interplay of cryptographic validation, behavioral verification, and adaptive trust management. Our results demonstrate that secure, verifiable, and privacy-preserving decentralized semantic learning is practically achievable at scale for intelligent transportation systems, paving the way for safer and more trustworthy autonomous collaboration in adversarial environments.