A two-stage framework for diffusion source localization via sensor-guided network pruning and temporal-topological alignment

Diffusion Source Localization (DSL) is essential for information security and network forensics, enabling the identification of origins for malicious cascades such as rumors, misinformation, and malware outbreaks. Existing methods face significant practical limitations due to their dependence on detailed diffusion information, including explicit transmission pathways or predefined propagation parameters. We propose Temporal-topological Region-constrained Approach for diffusion sourCe Estimation (TRACE), a lightweight DSL framework that operates effectively under sparse observation conditions by requiring only infection timestamps from a limited set of sensor nodes. TRACE employs a two-stage methodology to address both data scarcity and computational scalability. In the first stage, a Bounded Attempt Strategy (BAS) grounded in percolation theory optimizes sensor placement at structurally critical locations, enabling a simple topological partitioning procedure to isolate a compact candidate region containing the diffusion source. This stage significantly reduces the search space from (Formula presented) to (Formula presented) with c ≪ n. In the second stage, a non-parametric cosine-similarity-based estimator identifies the source within this reduced region by measuring the alignment between observed temporal patterns and topology-induced spatial distances. Comprehensive experiments on twelve real-world networks spanning social, communication, transportation, and infrastructure domains demonstrate that TRACE consistently outperforms state-of-the-art methods in localization accuracy, spatial precision, and computational efficiency. Further analyses confirm the robustness of TRACE under high-variance diffusion conditions and its scalability to large-scale networks.