Generative Client-Side Gradient Leakage Attack Via Poisoning-Driven Analytical Feature Separation in Federated Learning

Federated Learning (FL) has emerged as an important paradigm for distributed intelligence in large-scale AI-based IoT systems under the Integrated Sensing, Memory, Communication, and Computation (SMCC) framework. However, its privacy guarantees may be threatened by gradient leakage attacks under dynamic training conditions. Existing client-side attacks often suffer from limited reconstruction fidelity and high computational overhead. In this paper, we propose PAFS, a generative gradient leakage framework based on Poisoning-driven Analytical Feature Separation, designed to investigate privacy vulnerabilities relevant to SMCC-enabled collaborative learning. PAFS introduces a feature separation mechanism in which strategically poisoned updates amplify target-class gradients, enabling target-related features to be extracted from fully connected (FC) layers. To bridge the semantic–spatial gap in reconstruction, we further introduce a Hierarchical Feature Fidelity (HFF) mechanism that enforces multi-scale feature consistency to preserve both fine-grained textures and semantic structure. As a result, target images can be reconstructed via a single forward pass. Experiments on CIFAR100 and ImageNet show that PAFS achieves competitive reconstruction quality with improved efficiency compared with representative baselines, while maintaining stable performance under varying client scales and Byzantine-robust aggregation rules, and remaining effective under low-magnitude gradient perturbations. The results highlight potential privacy risks in SMCC-enabled FL systems and provide insights for privacy evaluation in intelligent IoT systems.