Skip to content

RATF: Robust Adaptive Tri-Encoder Framework for Fake News Detection against Style-Confusion Attacks from LLMs

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Iceldancer/RATF
Area: LLM (Other)
Keywords: Fake News Detection, Style-Confusion Attacks, Tri-Encoder Framework, Conditional Probability Curvature, Adaptive Gating

TL;DR

To resolve the fundamental tension where fake news detectors either fall prey to LLM-driven style-laundering attacks or sacrifice clean accuracy by enforcing rigid style-invariance, RATF decouples news into objective content and style descriptions via LLMs, computes an algorithmic rewritten score via conditional probability curvature, and adaptively calibrates tri-encoder representations through conditioned cross-attention and learnable parametric gating.

Background & Motivation

The rampant spread of fake news across digital and social media poses existential threats to political stability, public health, and societal cohesion. Conventional automated fake news detection models have historically derived strong discriminative power by combining semantic claims with stylistic and stylometric signals. Because unreliable sources and malicious fabricators frequently deploy sensationalist language, emotional appeals, and hyper-partisan framing in sharp contrast to the measured objectivity of reputable journalism, linguistic style has served as a highly reliable diagnostic marker. However, the advent and commoditization of advanced Large Language Models (LLMs) have fundamentally disrupted this foundation by industrializing text manipulation: bad actors can now trivially separate a fraudulent core claim from its original sensational tone and rewrite it to flawlessly mimic the measured, neutral register of prestigious journalistic institutions. This "style laundering" preserves falsehood while cloaking it in an aura of authoritative credibility.

By injecting these stylistically disguised articles into legitimate information streams alongside original unrewritten reports, adversaries construct what the authors formalize as a "style-confusion attack." Under this adversarial regime, conventional detectors encounter a crippling vulnerability: standard fine-tuned language models (such as RoBERTa and DeBERTa) experience precipitous performance declines as the proportion of style-attacked news rises. Conversely, existing defense paradigms centered on style-invariant representationsโ€”most notably Sheepdogโ€”deliberately strip away all stylistic cues to enforce invariant predictions across diverse rewritten variants. While this achieves robustness against style perturbation, it imposes a severe performance ceiling on authentic, unperturbed articles where stylistic indicators remain highly informative and legitimate. Fake news verification is thus caught in a debilitating dilemma: risking severe vulnerability by trusting linguistic style, or forfeiting critical detection efficacy by blindly discarding it.

To resolve this conflict, the authors contend that stylistic authenticity should neither be naively accepted nor entirely dismissed, but treated as a continuously estimable latent variable. By evaluating the likelihood that an article has been algorithmically rewritten, a verification architecture can dynamically recalibrate its reliance on style: leaning on rich stylistic patterns when the text is identified as authentic human writing, and adaptively dampening the style channel to anchor decisions strictly on objective factual substance when synthetic manipulation is detected. Core idea: decouple raw articles into objective content representations and explicit style features using an LLM, quantify the machine-rewriting likelihood via token conditional probability curvature, and dynamically calibrate semantic, content, and style representations through rewritten-score-conditioned cross-attention and learnable parametric gating.

Method

Overall Architecture

The RATF operational pipeline comprises three coordinated stages: (1) LLM-powered feature engineering and algorithmic rewriting detection, which decomposes the raw article \(A\) into objective content features \(A_{\text{con}}\), an explicit stylistic description \(A_{\text{sty}}\), and a continuous rewritten score \(s_{\text{rw}} \in [0, 1]\) via conditional probability curvature; (2) a Tri-Encoder module employing three independent DeBERTa-v3-base backbones to extract pooled vectors and token-level representations for original semantics, content, and style streams; and (3) a score-driven fusion and calibration network where the reliable content representation acts as a query to retrieve cross-modal evidence under score conditioning, while a learnable parametric gate dynamically attenuates potentially compromised streams prior to final classification.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    IN["Input News Article A<br/>(Real/Fake, Clean/Rewritten Mix)"] --> FE["LLM Dual-Stream Decoupling & Curvature Rewritten Detection<br/>Qwen2.5 distills content/style + GPT-Neo curvature gives srw"]
    FE --> TRI["Tri-Encoder Representation Extraction<br/>DeBERTa-v3 independently encodes original A, content Acon, style Asty"]
    TRI --> CAF["Score-Conditioned Cross-Attention Fusion<br/>pcon biased by srw serves as Query to attend Hsem and Hsty"]
    CAF --> GATE["Adaptive Learnable Parametric Score Gating<br/>End-to-end Sigmoid decay curves modulate wsem and wsty"]
    GATE --> OUT["Feature Concatenation & Veracity Classification<br/>Cross-Entropy optimization yielding final truth prediction"]

Key Designs

1. LLM Dual-Stream Decoupling & Curvature Rewritten Detection: Adversarial style-laundering obscures stylistic veracity -> Prompting LLM for factual/style separation and estimating conditional curvature -> Establishes unpolluted factual anchors and dynamic rewriting likelihood

To circumvent the rigid limitations of static dictionaries and handcrafted heuristics, RATF utilizes a general-purpose instruction-tuned LLM (Qwen2.5-7B-Instruct) via a single multi-part prompt to extract raw news into two explicit text streams: an objective rewritten text \(A_{\text{con}}\) representing the core factual substance purged of sensationalist tone, and a descriptive summary \(A_{\text{sty}}\) cataloging the original stylistic elements. Concurrently, to determine whether the incoming text has undergone algorithmic laundering, RATF builds upon the distributional characteristics of machine text identified in Fast-DetectGPT. Given an input text \(x\) and an autoregressive LM \(p_\theta\) (GPT-Neo), human text exhibits high conditional entropy and varied vocabulary choices, whereas LLMs preferentially sample predictable, high-probability tokens. For each position \(i\), \(N\) alternative candidate tokens are sampled from \(p_\theta(\cdot|x_{<i})\) to construct perturbation sequences whose mean \(\tilde{\mu}\) and variance \(\tilde{\sigma}^2\) of log-probabilities are calculated:

\[\tilde{\mu} = \frac{1}{N} \sum_{j=1}^N \log p_\theta(\tilde{x}^{(j)}), \quad \tilde{\sigma}^2 = \frac{1}{N} \sum_{j=1}^N \left(\log p_\theta(\tilde{x}^{(j)}) - \tilde{\mu}\right)^2\]

The conditional probability curvature \(d(x, p_\theta) = (\log p_\theta(x) - \tilde{\mu}) / \tilde{\sigma}\) measures the standardized z-score deviation. This metric is normalized and clipped into a probability-like score \(s_{\text{rw}} \in [0, 1]\), serving as an unsupervisable, model-agnostic barometer of algorithmic rewriting.

2. Score-Conditioned Cross-Attention Fusion: Heterogeneous feature streams suffer noise propagation -> Content representation guided by rewritten-score bias queries semantic and stylistic hidden states -> Factual-first cross-modal evidence aggregation

Given the three distinct textual inputs (\(A, A_{\text{con}}, A_{\text{sty}}\)), three parallel DeBERTa-v3-base encoders (\(E_{\text{sem}}, E_{\text{con}}, E_{\text{sty}}\)) generate pooled vectors \(h_{\text{sem}}, h_{\text{con}}, h_{\text{sty}} \in \mathbb{R}^{d_h}\) and token hidden states \(H_{\text{sem}}, H_{\text{sty}}\). After linear projection to \(p_{\text{sem}}, p_{\text{con}}, p_{\text{sty}}\), the objective content feature \(p_{\text{con}}\) is designated as the primary reference query because its factual essence remains intact under style perturbations. To ensure that the feature retrieval process is cognizant of the manipulation likelihood, the scalar rewritten score \(s_{\text{rw}}\) is transformed via a lightweight MLP \(\phi(\cdot)\) into a dimensional bias vector and injected directly into the content query:

\[q = \text{LayerNorm}(p_{\text{con}} + \phi(s_{\text{rw}}))\]

The conditioned query \(q\) then performs multi-head cross-attention across the token-level hidden states of original semantics \(H_{\text{sem}}\) and writing style \(H_{\text{sty}}\), followed by residual connections:

\[h_{\text{sem}}^{\text{ca}} = \text{MHA}(q, H_{\text{sem}}, H_{\text{sem}}), \quad h_{\text{sty}}^{\text{ca}} = \text{MHA}(q, H_{\text{sty}}, H_{\text{sty}})$$ $$h_{\text{sem}}' = p_{\text{sem}} + h_{\text{sem}}^{\text{ca}}, \quad h_{\text{con}}' = p_{\text{con}}, \quad h_{\text{sty}}' = p_{\text{sty}} + h_{\text{sty}}^{\text{ca}}\]

This query formulation ensures that content representations retrieve relevant contextual evidence while discounting deceptive semantic and stylistic artifacts in proportion to the rewriting probability.

3. Adaptive Learnable Parametric Score Gating: Hard invariant filtering penalizes clean data accuracy -> Continuous parametric Sigmoid attenuation dynamically re-weights style and semantic contributions -> Balanced adversarial robustness and clean performance preservation

Rather than applying a rigid binary threshold or uniformly zeroing out stylistic features, RATF implements an end-to-end learnable parametric gating mechanism driven solely by \(s_{\text{rw}}\). Because the objective content stream is verified to be reliable and fact-centered, its gating weight is perpetually maintained at \(w_{\text{con}}(s_{\text{rw}}) = 1\). For semantic and stylistic streams, adaptive Sigmoid attenuation curves modulate their contributions:

\[w_{\text{sem}}(s_{\text{rw}}) = \sigma\left(-k_{\text{sem}}(s_{\text{rw}} - b_{\text{sem}})\right), \quad w_{\text{sty}}(s_{\text{rw}}) = \sigma\left(-k_{\text{sty}}(s_{\text{rw}} - b_{\text{sty}})\right)\]

Here, \(k_{\text{sem}}, k_{\text{sty}} > 0\) control the decay steepness, and \(b_{\text{sem}}, b_{\text{sty}} \in (0, 1)\) specify the inflection centers, initialized at \(k_{\text{sem}}=4.0, b_{\text{sem}}=0.6\) and \(k_{\text{sty}}=15.0, b_{\text{sty}}=0.5\) respectively. Both parameters are optimized jointly with the downstream classification loss. Consequently, when \(s_{\text{rw}} \to 0\) (genuine human writing), style and semantic weights remain near 1.0, unleashing full stylometric discriminability; conversely, when \(s_{\text{rw}} \to 1\) (high likelihood of LLM rewriting), \(w_{\text{sty}}\) decays rapidly toward 0, insulating the classifier from deceptive style cues.

Loss & Training

The calibrated representations \(\hat{h}_{\text{sem}} = w_{\text{sem}} h_{\text{sem}}'\), \(\hat{h}_{\text{con}} = w_{\text{con}} h_{\text{con}}'\), and \(\hat{h}_{\text{sty}} = w_{\text{sty}} h_{\text{sty}}'\) are concatenated into \(h_{\text{merged}} = [\hat{h}_{\text{sem}}; \hat{h}_{\text{con}}; \hat{h}_{\text{sty}}]\). The merged vector is passed through a fully connected projection layer with LayerNorm and ReLU non-linearity to yield \(h_{\text{fused}}\), which is projected onto class logits \(z \in \mathbb{R}^2\). Model parameters are optimized end-to-end using standard Cross-Entropy loss:

\[\mathcal{L}_{\text{CE}} = - \sum_{c=1}^2 y_c \log \frac{\exp(z_c)}{\sum_{j=1}^2 \exp(z_j)}\]

The model is trained using the AdamW optimizer with a learning rate of \(2 \times 10^{-5}\), a weight decay of 0.01, a batch size of 32, and a linear learning rate scheduler with zero warmup steps, selecting the best-performing checkpoint on the validation split for testing.

Key Experimental Results

Main Results

Evaluation was conducted on PolitiFact and GossipCop from FakeNewsNet following temporal split protocols. The attack scenario introduces four stylistic rewriting variants (sensational, objective, partisan, and general style transfer) downsampled into a 1:1 balanced mix of clean and rewritten news with 50:50 real/fake distributions. Baselines span zero-shot LLMs (G1: Gemini 2.5 Flash, GPT5.4), fine-tuned SLMs (G2: RoBERTa, DeBERTa, PET), and LLM-enhanced verification systems (G3: Sheepdog, ARG, TED).

Dataset Group / Method Accuracy Weighted-F1 Precision Recall
PolitiFact (confusion) G1: Gemini 2.5 Flash 0.6638 0.6852 0.6476 0.7083
PolitiFact (confusion) G1: GPT5.4 0.7149 0.7124 0.7155 0.7129
PolitiFact (confusion) G2: RoBERTa 0.7528 0.7634 0.7030 0.8353
PolitiFact (confusion) G2: DeBERTa 0.7921 0.7702 0.8158 0.7249
PolitiFact (confusion) G2: PET 0.7536 0.7651 0.7124 0.7847
PolitiFact (confusion) G3: Sheepdog 0.8258 0.8238 0.8380 0.8258
PolitiFact (confusion) G3: ARG 0.8467 0.8465 0.8475 0.8464
PolitiFact (confusion) G3: TED 0.8378 0.8379 0.8382 0.8376
PolitiFact (confusion) Ours (RATF) 0.8820 0.8823 0.8820 0.8819
GossipCop (confusion) G1: Gemini 2.5 Flash 0.6039 0.6023 0.5803 0.6192
GossipCop (confusion) G1: GPT5.4 0.6353 0.6261 0.6292 0.6337
GossipCop (confusion) G2: RoBERTa 0.7361 0.7409 0.7207 0.7621
GossipCop (confusion) G2: DeBERTa 0.7554 0.7523 0.7540 0.7506
GossipCop (confusion) G2: PET 0.7374 0.7362 0.7343 0.7385
GossipCop (confusion) G3: Sheepdog 0.7749 0.7747 0.7753 0.7748
GossipCop (confusion) G3: ARG 0.7456 0.7657 0.7726 0.7641
GossipCop (confusion) G3: TED 0.7538 0.7540 0.7659 0.7636
GossipCop (confusion) Ours (RATF) 0.7813 0.7816 0.7811 0.7807

On unperturbed original test sets (Table 5), RATF further confirms its superiority over all baselines: PolitiFact Acc/F1 achieves 92.17% / 92.26% (surpassing Sheepdog's 88.44%/88.39% and TED's 90.25%/90.37%), while GossipCop reaches 76.51% / 76.51% (outperforming Sheepdog's 75.77%/75.75% and TED's 76.12%/76.17%).

Ablation Study

The modular ablation isolates the specific utility of the rewritten score, conditioned attention, and parametric gating across both benchmarks:

Model Configuration PolitiFact Acc PolitiFact F1 GossipCop Acc GossipCop F1 Description / Findings
Ours (RATF Full Model) 0.8820 0.8823 0.7813 0.7816 Full tri-encoder + SC-CAF + Learnable Gating
RATF w/o Rewritten Score 0.8487 0.8464 0.7526 0.7524 Removes \(s_{\text{rw}}\) signal from attention and gating
SC-CAF Only 0.8743 0.8754 0.7566 0.7561 Preserves conditioned cross-attention, removes gating
Learnable Gate Only 0.8516 0.8507 0.7604 0.7615 Preserves parametric gate, removes SC-CAF interaction
Content-Only 0.8315 0.8308 0.7431 0.7429 Single-stream baseline using only \(A_{\text{con}}\) features
Style-Only 0.7753 0.7744 0.7052 0.7044 Single-stream baseline using only \(A_{\text{sty}}\) features

Backbone adaptability experiments (Table 4) demonstrate substantial, stable gains across base architectures: RoBERTa improves from 76.34% to 87.85% (+11.51%) on PolitiFact and from 74.09% to 77.94% (+3.85%) on GossipCop; BERT improves from 75.36%/73.11% to 86.31%/75.68%; DeBERTa improves from 77.02%/75.23% to 86.72%/77.89%, illustrating model-agnostic utility.

Key Findings

  • The rewritten score is the indispensable control signal: Stripping the continuous curvature score (w/o Rewritten Score) precipitates a steep drop in F1 of 3.59% on PolitiFact and 2.92% on GossipCop, confirming that unguided attention fusion cannot distinguish between reliable and manipulated textual evidence.
  • Cross-attention and parametric gating serve complementary roles: SC-CAF excels at token-level evidence retrieval (achieving 0.8754 F1 on PolitiFact), while Learnable Gate provides macro-level attenuation against stylistic noise. Employing both mechanisms yields optimal performance across both political and celebrity news.
  • Content provides a dependable floor while style unlocks the performance ceiling: The Content-Only model achieves competitive standalone performance (0.8308 / 0.7429 F1), validating the quality of Qwen2.5 feature extraction. Conversely, Style-Only degrades to 0.7744 / 0.7044 under attack. The full model outperforms Content-Only by over 5% on PolitiFact, demonstrating that properly calibrated style features offer vital positive utility.
  • Domain sensitivity differs across claim lengths: Gains on PolitiFact are substantially larger than on GossipCop. Political claims are longer and heavily saturated with partisan and sensational rhetoric, rendering style shifts more pronounced, whereas entertainment gossip is structurally concise and formulaic.

Highlights & Insights

  • Soft adaptive recalibration over rigid binary choices: Rather than falling into the trap of either naively trusting linguistic style or rigidly enforcing style-invariance, RATF reframes style utility into a continuous function of algorithmic manipulation likelihood, establishing a nuanced defense paradigm.
  • Zero-shot curvature estimation for robust attribution: Adapting conditional probability curvature from Fast-DetectGPT enables RATF to reliably measure machine-generated likelihood without fine-tuning specialized neural discriminators, conferring resilience against evasion attempts across varied LLM generators.
  • Decoupled execution and backbone neutrality: Feature extraction requires only a single-pass prompt without recursive multi-agent debates, and the downstream tri-encoder seamlessly accommodates BERT, RoBERTa, or DeBERTa architectures.

Limitations & Future Work

  • Inference latency from external LLM components: Computing curvature scores and extracting dual text streams via Qwen2.5-7B and GPT-Neo introduces additional inference overhead compared to bare small encoders, which may limit applicability in ultra-high-throughput real-time streaming settings.
  • Evaluation breadth limited to benchmark distributions: The empirical framework was evaluated on PolitiFact and GossipCop; testing across multilingual datasets, non-Western political environments, and complex multimodal news streams remains to be validated.
  • Absence of open-world factual retrieval: Predictions are derived internally from text distributions without querying live external knowledge graphs or web corpora, suggesting future integration with Retrieval-Augmented Generation (RAG) to counter deeply fabricated claims.
  • vs Sheepdog (KDD 2024): Sheepdog trains small detectors to ignore stylistic variations via multi-style reframing and adversarial invariant objectives. While resistant to style shifts, it incurs an unavoidable penalty on clean news; RATF dynamically estimates style authenticity, retaining full stylometric benefits on unattacked articles while remaining resilient under attack.
  • vs ARG (AAAI 2024) & TED (SIGIR 2025): ARG integrates multi-perspective LLM rationales and TED stages proponent-opponent agent debates. However, both frameworks can be deceived when false claims are cloaked in authoritative prose; RATF directly detects low-level probability curvature and anchors retrieval to an objective factual stream, achieving higher robustness at lower interactive computational cost.

Rating

  • Novelty: โญโญโญโญโญ Formulates the style-confusion attack and introduces a continuous curvature-driven gating mechanism over tri-encoder representations.
  • Experimental Thoroughness: โญโญโญโญโญ Rigorous comparisons across 8 baselines spanning 3 model families, exhaustive ablations, clean data evaluations, and backbone generalizability tests.
  • Writing Quality: โญโญโญโญโญ Clear motivational narrative, precise mathematical formulation of curvature and gating curves, and well-structured empirical analyses.
  • Value: โญโญโญโญโญ Provides an effective and deployable blueprint for safeguarding automated fact-checking systems against LLM-powered style laundering.