Detailed Cryptanalysis of “Privacy-Preserving Quantum Federated Learning via Gradient Hiding”

Quantum Federated Learning (QFL) has crystallized as a formidable paradigm that aspires to reconcile distributed intelligence with uncompromised privacy. Notably, the protocol articulated by Changhao Li et al. [1], which harnesses gradient concealment through Blind Quantum Bipartite Correlators and GHZ-entangled states, epitomizes this ambition. However, its ostensible guarantees remain largely untested against sophisticated adversarial models capable of subverting quantum safeguards. In this study, we undertake a rigorous cryptographic dissection of the protocol, constructing formal threat models encompassing twelve distinct attack vectors, from phase manipulation and amplitude distortion to entanglement erosion and replay amplification. Through meticulous mathematical scrutiny, we demonstrate that these vectors can precipitate the disclosure of private gradients and compromise aggregation fidelity. To remediate these deficiencies, we propose a cohesive set of countermeasures, including randomized phase obfuscation, authenticated quantum encodings, and temporal binding strategies. These contributions collectively advance a resilient architectural framework, charting a credible path toward quantum-secure federated learning in adversarial settings.