Validator nodes in blockchain networks face network-layer threats in which traditional enterprise firewalls do not adequately address. In this work, we introduce BCFW, a Blockchain Consensus Firewall that bridges flow-based intrusion detection with on-chain governance for Proof-of-Authority (PoA) deployment. At its core, an FT-Transformer processes the 41 KDD Cup 99 features as tokens, using self-attention to model cross-feature interactions and outputting multi-class predictions with calibrated confidence. Class imbalance is mitigated through log-smoothed weighting and label smoothing. Upon threat detection, an orchestration layer may submit proposed mitigation measures to a multisignature contract. Validators then conduct an on-chain vote, and approved responses—such as rate limiting, isolation, or access-control updates—are automatically executed with complete audit trails. On the KDD Cup 99 benchmark, BCFW achieves 99.95% accuracy and 0.9182 macro-F1 across five attack categories. Our results show that coupling neural detection with blockchain-native governance yields an auditable, collectively controlled defense pipeline for validator infrastructure.