QFedPhish: A Quantum-Inspired Federated Learning Approach for Detecting AI-Generated Phishing Attacks

Advancements in artificial intelligence (AI) offer opportunities but have also contributed to the emergence of AIgenerated attacks, escalating the complexity and scale of modern cyber threats. These attacks generate serious concerns for the traditional security systems in maintaining accuracy, privacy, false-positive control, and resilience in modern distributed environments. To cope with these concerns, this paper proposes QFedPhish, a quantum-inspired federated learning approach designed to identify and detect AI-generated phishing attacks. A combination of real-world phishing logs with synthetically generated large language model-based phishing attempts is built as a representative dataset. This data is then transformed through a quantum-inspired feature encoder, where random Fourier features and tensor-network compression capture subtle stylometric and structural cues that distinguish phishing content from legitimate content. Next, each participating client locally trains a lightweight classifier on the resulting embeddings, enabling model adaptation to heterogeneous environments without exposing raw data. Finally, model updates are aggregated using federated learning enhanced with differential privacy and resilient aggregation mechanisms. It mitigates risks of data leakage and poisoning attacks. Experimental evaluation demonstrates that QFedPhish consistently outperforms the best benchmark models, achieving gains of 0.10% in accuracy, 0.33% in precision, 0.41% in F1-score, 0.61% in AUC, along with reducing communication overhead by 3.89% and improving convergence speed by 3.66%.