Hiding the Rumor Source

Anonymous social media platforms like Secret, Yik Yak, and Whisper have emerged as important tools for sharing ideas without the fear of judgment. Such anonymous platforms are also important in nations under authoritarian rule, where freedom of expression and the personal safety of message authors may depend on anonymity. Whether for fear of judgment or retribution, it is sometimes crucial to hide the identities of users who post sensitive messages. In this paper, we consider a global adversary who wishes to identify the author of a message; it observes either a snapshot of the spread of a message at a certain time, sampled timestamp metadata, or both. Recent advances in rumor source detection show that existing messaging protocols are vulnerable against such an adversary. We introduce a novel messaging protocol, which we call adaptive diffusion, and show that under the snapshot adversarial model, adaptive diffusion spreads content fast and achieves perfect obfuscation of the source when the underlying contact network is an infinite regular tree. That is, all users with the message are nearly equally likely to have been the origin of the message. When the contact network is an irregular tree, we characterize the probability of maximum likelihood detection by proving a concentration result over Galton-Watson trees. Experiments on a sampled Facebook network demonstrate that adaptive diffusion effectively hides the location of the source even when the graph is finite, irregular and has cycles.

💡 Research Summary

The paper addresses the problem of protecting the identity of a user who originates a message on an anonymous social‑media platform when a powerful global adversary can observe either a snapshot of the infected nodes, timestamps collected from compromised (“spy”) nodes, or both. Existing rumor‑source detection methods show that the standard diffusion model—where each infected node independently forwards the message to its neighbors with a fixed probability—allows such an adversary to locate the source with high probability using graph‑centrality measures (rumor centrality, Jordan center, etc.).

To counter this vulnerability, the authors propose Adaptive Diffusion, a novel spreading protocol that deliberately breaks the symmetry of ordinary diffusion. The key idea is to maintain a “virtual source” that moves from node to node in odd time steps; in the following even step the protocol expands a perfectly balanced subtree rooted at the new virtual source. By inserting artificial delays, the protocol ensures that at every even time (t) the infected subgraph (G_t) is a depth‑(t/2) regular tree (or a balanced tree on a (d)-regular infinite graph). Consequently, all infected nodes lie at the same distance from the virtual source and have identical structural positions.

On an infinite (d)-regular tree this construction yields perfect obfuscation: given only the set of infected nodes, the maximum‑likelihood (ML) estimator assigns equal probability (1/|V_T|) to each infected node, so the probability of correctly identifying the true source is exactly the reciprocal of the infected population. The authors prove this result formally and also derive a matching lower bound, showing that no protocol can achieve a lower detection probability under the snapshot adversary model.

For more realistic, irregular trees the authors analyze a family of Galton‑Watson trees. They prove a concentration result for the “extreme paths” (paths containing high‑degree nodes) and show that the detection probability under Adaptive Diffusion still converges to a constant strictly less than one, depending only on the offspring distribution’s mean and variance. This demonstrates that the protocol remains robust even when the perfect symmetry of a regular tree cannot be maintained.

Empirical evaluation is performed on a sampled Facebook subgraph and on several synthetic networks (Erdős‑Rényi, random regular, lattice). The results confirm that Adaptive Diffusion spreads almost as fast as the optimal diffusion scheme—its time to reach (n) nodes is at most twice the optimal—but achieves dramatically lower detection rates (often below 10 %). The protocol also performs well against the spy‑based and combined spy+snapshot adversaries, as discussed in the companion technical report