Detecting Camouflaged Rumors from Where Replies Disagree

Language models can now rewrite a false claim so it reads like a calm wire-service report. Detectors that learned to spot fakes from sensational wording lose up to 38% F1 when this happens (Fake News in Sheep's Clothing). During breaking news there is also often nothing to check the claim against, because encyclopedias, search indexes and fact-check reports lag the event (XFacta; MMFakeBench). In that window the main surviving evidence is how people respond: who questions the post, who denies it, and whether those challenges get answered. Current systems read this reply signal by averaging reply features over the conversation tree. They either count stances or run a standard graph neural network, meaning a model that repeatedly mixes each node's features with its neighbors'. Averaging is exactly what an adversary can defeat cheaply. Flooding a thread with fluent, supportive replies dilutes any average.

We propose that the veracity signal that survives both style camouflage and missing knowledge lies in local disagreement across reply edges, not in the thread-wide consensus. Concretely, a child reply may contradict its parent, a skeptic may sit adjacent to the root, or a query may go unanswered. We build an interpretable detector whose intermediate "concepts" are these edge-level disagreement quantities. It predicts rumor veracity through a sparse, human-readable layer. We test whether it keeps its accuracy when both the source post and the replies are camouflaged by a language model and no external evidence is available. The two-sentence version: averaging replies is fragile because injected replies dilute it, while localized contradictions are not diluted, so build the explanation layer out of contradictions.

Core Claim

Graph over-smoothing is usually discussed as a depth problem. Here it serves as a lens on an attack surface. A mean-aggregating graph layer acts as a low-pass filter: it keeps what neighbors share and discards how they differ (Complete the Missing Half). Let $s_v$ be a reply's predicted stance distribution over support, deny, query and comment. Let $P$ be the row-normalized parent–child adjacency of the conversation tree. The low-pass signal is $Ps$ and the high-pass signal is

$$h = (I - P)\,s,$$

which is nonzero only where a reply differs from its conversational neighbors.

Consider an attacker who adds $k$ supportive replies to a thread of $n$ genuine replies. That attacker shifts any thread-wide stance fraction by a factor of roughly $n/(n+k)$, since the statistic is a mean. A statistic built from the largest or top-$m$ entries of $h$ near genuine skeptics does not move unless the attacker replies directly to those skeptics. Our claim is threefold. First, concepts computed from $h$ give detection that degrades far more slowly under combined source-rewriting and reply-injection attacks than stance counts or mean-aggregating graph networks. Second, this holds with the external evidence channel removed entirely. Third, it costs little or no clean accuracy relative to opaque graph models.

This differs from each close neighbor. LOSIRD (Meet The Truth) pairs a Wikipedia evidence graph with a conversation tree encoded by GraphSAGE. Its interpretability comes from surfaced evidence sentences, which disappear when retrieval returns "not enough information." Its conversation channel is a low-pass averager, and it was never tested against camouflaged text. SheepDog (Fake News in Sheep's Clothing) makes a text detector style-invariant and adds LLM-generated veracity attributions as flat auxiliary labels. It reads only the article, so it has nothing to fall back on when the article itself is perfectly camouflaged. InSide (Bridging Thoughts and Words) builds intent and semantic graphs inside the article, not over the audience's reaction. Graph concept bottlenecks (Graph Concept Bottleneck Models) pass messages among concepts in vision and medical tasks. We instead define the concepts themselves as graph-filter outputs, and we keep the concept-to-label head linear so no smoothing happens after the bottleneck.

Research Program

The detector has three stages. The first is a per-reply stance encoder, a fine-tuned text model. It is trained with style-reframing augmentation in the manner of SheepDog, so that a polite denial and a hostile denial map to the same stance. Stance supervision comes from an LLM teacher, following SheepDog's use of LLMs as offline teachers rather than zero-shot judges. The teacher is audited on a hand-annotated sample of threads. The second stage is a fixed, parameter-free computation of concepts from the stance field on the tree. The concepts are the maximum and top-$m$ edge disagreement at depth one (challenges to the root), and the same quantities at deeper levels. They also include the fraction of query or deny replies that receive an evidential rebuttal, and the timing of the first root-adjacent denial. As a deliberate contrast, we also compute the conventional low-pass concepts, which are thread-wide stance fractions. The third stage is a sparse logistic head over the concepts, so each prediction decomposes into named, inspectable contributions. It also supports test-time intervention in the sense of Concept Bottleneck Models: an analyst can correct a mis-tagged stance and see the prediction update.

Three analyses turn this from an architecture into a finding. The first is a spectral split. We train otherwise identical heads on low-pass-only, high-pass-only, and combined concepts, and we measure how much veracity information each carries on clean threads and under attack. This tests directly whether the robust signal lives where we claim.

The second is a dilution curve. We inject $k$ LLM-written replies, styled to sound credible, at random positions, and we plot accuracy against $k/n$. The mean-based theory predicts a specific decay shape for count concepts and near-flatness for edge concepts. An adaptive attacker who targets skeptics directly is the stress test.

The third addresses the distraction effect. Veltmeijer et al. found that forcing clinical concepts can hurt some classes (Integrating Clinically-Relevant Features into Skin Lesion Classification), whereas SheepDog reports that auxiliary attributions help. We measure per-class cost of the bottleneck against an unconstrained model with the same inputs to see which result holds for disagreement concepts.

Baselines

The naive reference points come first. One is a majority-class predictor. The other is a text-only RoBERTa classifier on the source post. Under source rewriting we expect it to fall toward the collapse SheepDog documents. If our method fails to beat it clearly under attack, the reply channel adds nothing.

The strongest existing method on conversation threads is LOSIRD (Meet The Truth). On PHEME 2017 and 2018 it reports 91.4% and 92.5% accuracy with its Wikipedia evidence graph included. We run it twice: with evidence as published, and with the evidence graph forced to "not enough information" to simulate zero-day latency. We found no reported result on Lacuna for LOSIRD without evidence or under stylistic attack, so we produce those numbers ourselves with the released design, identical splits and an equal tuning budget. Beating it means matching its clean zero-evidence accuracy within about one point while losing substantially less under combined attacks.

The strongest text-side competitors are SheepDog and AdStyle (Adversarial Style Augmentation via Large Language Model for Robust Fake News Detection). AdStyle reports an AUC of 0.9174 on PolitiFact under attack, against 0.8199 for SheepDog. That number comes from a different dataset without reply threads, so it motivates but cannot directly anchor our comparison. We retrain both on PHEME source posts with the same LLM budget used to build our attacks.

Two further baselines isolate our key ingredients. A two-channel filterbank graph network (Complete the Missing Half) on the same conversation tree tests whether high-pass information alone, without a bottleneck, explains the robustness. If it matches us, the concept layer is purely an interpretability benefit. An entropy-gated aggregator (When Do GNNs Work) gives a second opaque anti-smoothing competitor. Finally, the stance-count concept model is our own architecture with only the low-pass concepts. It is the ablation that isolates the edge-disagreement idea, and the proposal stands or falls on beating it under reply injection.

Why not just…?

Why not just use a two-layer GraphSAGE; aren't reply trees too shallow to over-smooth? Depth is not the issue. The issue is that even one mean-aggregation step blends a skeptic with the replies around it, and filterbank results show low-pass-only aggregation fails on disagreeing neighborhoods even when shallow (Complete the Missing Half). Rumor threads may nonetheless be homophilic enough that this does not matter on clean data. The dilution curve settles it: a shallow averager should decay with injected replies the way count concepts do.

Why not just count denying and questioning replies with logistic regression? This is the most serious threat. The count model is simple, interpretable and style-invariant once stance tagging is robust, and it may match us on clean PHEME threads. Our argument is only that counts are means, and means dilute. If the count model degrades no faster than ours under reply injection, our core claim is false and we will report it as such. The stance-count ablation is built to decide exactly this.

Why not just harden the text detector with SheepDog or AdStyle? Style-invariant training helps when some content cue survives rewriting. But entity-swapped or fully neutral fabrications leave the text stylistically identical to real news (Automatic Detection of Entity-Manipulated Text using Factual Knowledge), and AdStyle's repeated LLM calls are costly. We include both as baselines. We also test a late-fusion variant, since hardened text and disagreement concepts may be complementary rather than competing.

Why not just prompt GPT-4o with the thread, or retrieve web evidence? Zero-evidence GPT-4o drops to 70.5% on post-cutoff events, and retrieval lifts it to 89.5% only when indexed evidence exists (XFacta). MMD-Agent degrades when Wikipedia has nothing on a breaking claim (MMFakeBench). We still run GPT-4o zero-shot on the full thread without retrieval as a baseline, because a frontier model reading replies might implicitly find disagreement itself.

Proposed Experiments

The main benchmark is PHEME, which consists of Twitter threads from breaking crises, with Weibo as a second-language replication. Both are split chronologically by event, so that test events are unseen, and both are evaluated with the evidence channel removed. Attacks come in three families. Source rewriting uses SheepDog-style reframings (objective, neutral, professional). Reply injection adds credible-sounding supportive replies at ratios $k/n$ from 0 to 1. The adaptive attack places rebuttals directly under genuine skeptics. We report accuracy and macro F1 on clean and attacked threads, the area under the dilution curve, and early-detection curves as a function of the first 5, 10, 20 and 50 replies to reflect the latency window. Interpretability is measured by intervention faithfulness, that is, how often correcting a single mis-tagged stance flips a wrong prediction to right, and by per-class cost relative to the unconstrained filterbank model.

The decisive checks are these. The high-pass-only head should retain most of the combined head's accuracy under reply injection while the low-pass-only head decays roughly in proportion to $n/(n+k)$. Our model should beat the stance-count ablation and zero-evidence LOSIRD under combined attacks, without losing more than about a point on clean data. Under the adaptive attack, we expect the injected rebuttals to register as more disagreement rather than less, which would make camouflage self-defeating. If instead that attack succeeds, it marks the boundary of the method, and we will report it as such.