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 Reasoning
Keywords: fake news detection, style-confusion attacks, tri-encoder architecture, conditional probability curvature, adaptive feature gating

TL;DR

To defend against LLM-driven style-confusion attacks where false claims mimic trusted journalistic tones within a mixed stream of authentic articles, RATF introduces a tri-encoder framework driven by conditional probability curvature rewritten scores, adaptively balancing objective content and stylistic signals via conditioned cross-attention and learnable parametric gating.

Background & Motivation

Automated fake news detection has long capitalized on joint representations of factual content and linguistic style, exploiting distinguishing patterns such as sensationalism, emotional charge, and partisan framing typical of fabricated reports. However, the rise of Large Language Models (LLMs) has industrialized misinformation generation, providing malicious actors with the tools to decouple inflammatory fake claims from their crude presentation and reconstruct them to mimic reputable, objective journalism. When attackers inject these stylistically disguised articles into streams of authentic news, conventional detectors (e.g., fine-tuned RoBERTa or DeBERTa) suffer catastrophic performance degradation because their learned stylistic priors are systematically subverted.

Existing defenses against such stylistic evasion predominantly pursue style-invariance, as exemplified by Sheepdog, which forces representations to disregard stylistic variance across multiple LLM-generated reframings. While discarding style renders these systems resilient to adversarial style shifts, it establishes an inherent performance ceiling: in clean, non-manipulated articles, genuine stylistic traits serve as indispensable discriminative cues. Forcing total style-invariance forfeits these natural signals, penalizing clean-data accuracy. Conversely, fully relying on stylistic features leaves detectors completely vulnerable to style manipulation.

Resolving this fundamental dilemma requires shifting away from the binary choice of either blindly trusting or completely ignoring style, treating stylistic authenticity as a continuous, estimable property. Because human writing and autoregressive LLM generations display distinct distributions in token-level conditional probability curvature, one can measure the likelihood of machine intervention without requiring supervision. Core idea: compute an unsupervised rewritten score via conditional probability curvature to continuously condition a tri-encoder framework—using objective content as an anchor in score-conditioned cross-attention and adaptively down-weighting potentially corrupted style signals through learnable parametric gating.

Method

Overall Architecture

RATF executes across four stages: LLM-driven feature decoupling, curvature-based rewritten score estimation, score-conditioned cross-attention fusion, and learnable parametric feature gating. First, an instruction-tuned LLM (Qwen2.5-7B-Instruct) decomposes each article in a single prompt into an objective factual rewrite (\(A_{\text{con}}\)) and a textual style description (\(A_{\text{sty}}\)). Simultaneously, an autoregressive language model (GPT-Neo) samples alternative tokens to compute the text's conditional probability curvature, yielding a normalized rewritten score \(s_{\text{rw}} \in [0, 1]\). Next, three parallel DeBERTa-v3-base encoders extract multi-view representations. The content vector acts as a robust query conditioned on \(s_{\text{rw}}\) to retrieve fine-grained tokens from the semantic and style streams. Finally, a learnable parametric gate continuously discounts vulnerable feature streams before a classification head outputs the veracity prediction.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    IN["Input News Article A"] --> D1["LLM-Driven Content and Style Disentanglement"]
    IN --> D2["Conditional Probability Curvature Rewritten Score Estimation"]
    D1 --> D3["Tri-Encoder Multi-View Representation Extraction"]
    D3 --> D4["Score-Conditioned Cross-Attention Fusion"]
    D2 -->|Rewritten Score srw| D4
    D4 --> D5["Learnable Parametric Feature Attenuation Gating"]
    D2 -->|Rewritten Score srw| D5
    D5 --> OUT["Concatenation and Fully-Connected Classification Head"]

Key Designs

1. LLM-Driven Content and Style Disentanglement: isolating factual substance from linguistic rhetoric without fragile manual heuristics

Traditional stylometric defenses depend on fixed lexicons, emotion dictionaries, or shallow syntactic hand-crafting that fail to separate factual claims from persuasive styling. RATF replaces manual heuristics by prompting Qwen2.5-7B-Instruct with a unified multi-part prompt to concurrently generate an objective descriptive version of the news (\(A_{\text{con}}\)) and an explicit textual description of the original stylistic elements (\(A_{\text{sty}}\)). Because \(A_{\text{con}}\) preserves the core factual assertions while stripping away sensational or deceptive rhetoric, it provides an uncorrupted anchor for subsequent veracity reasoning even when an adversary radically alters surface styling.

2. Conditional Probability Curvature Rewritten Score Estimation: unsupervised zero-shot assessment of machine rewriting likelihood

To empower the downstream detector to gauge the risk of stylistic manipulation, RATF computes the conditional probability curvature inspired by Fast-DetectGPT. The underlying hypothesis is that human authors select words with higher entropy and variance across alternative word choices, whereas autoregressive LLMs consistently gravitate toward high-probability token paths. Given an input text \(x\) and a causal language model \(p_\theta\) (GPT-Neo), for each position \(i\), \(N\) alternative tokens \(\tilde{x}_i\) are sampled from \(p_\theta(\cdot \mid x_{<i})\) to construct perturbed sequence candidates \(\{\tilde{x}^{(1)}, \dots, \tilde{x}^{(N)}\}\). The log-probabilities of these sequences yield mean \(\tilde{\mu}\) and variance \(\tilde{\sigma}^2\). The conditional probability curvature \(d(x, p_\theta)\) is defined as the standardized \(z\)-score:

\[d(x, p_\theta) = \frac{\log p_\theta(x) - \tilde{\mu}}{\tilde{\sigma}}\]

This metric is mapped into a normalized scalar \(s_{\text{rw}} \in [0, 1]\) via normalization and clipping. Operating without labeled detection data or fine-tuning, \(s_{\text{rw}}\) serves as a continuous, reliable indicator of algorithmic rewriting.

3. Tri-Encoder Multi-View Representation Extraction: independent representation spaces preventing cross-stream interference

Given the original text \(A\), objective content \(A_{\text{con}}\), and style descriptions \(A_{\text{sty}}\), passing heterogeneous signals into a single encoder causes cross-modal interference. RATF deploys three separate pre-trained deberta-v3-base backbones (\(E_{\text{sem}}, E_{\text{con}}, E_{\text{sty}}\)) to process the streams independently:

\[h_{\text{sem}} = E_{\text{sem}}(A), \quad h_{\text{con}} = E_{\text{con}}(A_{\text{con}}), \quad h_{\text{sty}} = E_{\text{sty}}(A_{\text{sty}})\]

The pooled outputs are projected via linear layers to dimension \(d_h\), yielding \(p_{\text{sem}}, p_{\text{con}}, p_{\text{sty}}\), while retaining token-level hidden states \(H_{\text{sem}}\) and \(H_{\text{sty}}\). This structure ensures that factual semantics and rhetorical style inhabit distinct latent manifolds.

4. Score-Conditioned Cross-Attention Fusion: content-anchored evidence retrieval conditioned on manipulation risk

Under style-confusion attacks, the content representation \(p_{\text{con}}\) provides the most uncorrupted factual core. RATF employs \(p_{\text{con}}\) as the primary query, adding an offset generated by feeding \(s_{\text{rw}}\) through a lightweight MLP \(\phi(\cdot)\) to form a score-conditioned query:

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

Using \(q\), the model queries the token-level semantic and stylistic sequences via multi-head cross-attention:

\[h^{\text{sem}}_{\text{ca}} = \text{MHA}(q, H_{\text{sem}}, H_{\text{sem}}), \quad h^{\text{sty}}_{\text{ca}} = \text{MHA}(q, H_{\text{sty}}, H_{\text{sty}})\]

Residual connections update the representations: \(\tilde{h}_{\text{sem}} = p_{\text{sem}} + h^{\text{sem}}_{\text{ca}}\), \(\tilde{h}_{\text{sty}} = p_{\text{sty}} + h^{\text{sty}}_{\text{ca}}\), and \(\tilde{h}_{\text{con}} = p_{\text{con}}\). This enables factual queries to selectively absorb supporting evidence while adjusting retrieval patterns according to rewriting severity.

5. Learnable Parametric Feature Attenuation Gating: end-to-end continuous calibration balancing clean accuracy and adversarial resilience

To avoid static thresholds or hard feature pruning, RATF introduces learnable parametric Sigmoid attenuation curves driven exclusively by \(s_{\text{rw}}\). The objective content weight is fixed at \(w_{\text{con}}(s_{\text{rw}}) = 1\), while semantic and stylistic gates follow parameterized logistic decays:

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

Here, slopes \(k_{\text{sem}}, k_{\text{sty}} > 0\) control transition sharpness, and centers \(b_{\text{sem}}, b_{\text{sty}} \in (0, 1)\) govern decay thresholds. When \(s_{\text{rw}}\) is near zero (authentic human article), gate values remain close to 1, maximizing stylistic utility. When \(s_{\text{rw}}\) approaches 1 (adversarial rewrite detected), the style and original semantic paths are aggressively attenuated, compelling the model to base predictions strictly on objective factual content.

Loss & Training

The calibrated vectors \(\hat{h}_{\text{sem}} = w_{\text{sem}}\tilde{h}_{\text{sem}}\), \(\hat{h}_{\text{con}} = w_{\text{con}}\tilde{h}_{\text{con}}\), and \(\hat{h}_{\text{sty}} = w_{\text{sty}}\tilde{h}_{\text{sty}}\) are concatenated into \(h_{\text{merged}} = [\hat{h}_{\text{sem}}; \hat{h}_{\text{con}}; \hat{h}_{\text{sty}}]\) and passed through a fully-connected layer with LayerNorm and ReLU. The output logits \(z \in \mathbb{R}^2\) are optimized via standard cross-entropy loss:

\[\mathcal{L}_{\text{CE}} = - \sum_{c \in \{0, 1\}} y_c \log \left(\frac{\exp(z_c)}{\sum_{j=0}^1 \exp(z_j)}\right)\]

The gating parameters are initialized to \(k_{\text{sem}} = 4.0, b_{\text{sem}} = 0.6\) and \(k_{\text{sty}} = 15.0, b_{\text{sty}} = 0.5\) (reflecting that stylistic features require steeper attenuation once manipulation is suspected) and optimized end-to-end. Training is carried out using AdamW with a learning rate of \(2 \times 10^{-5}\), weight decay of 0.01, batch size of 32, and linear learning rate decay without warmup.

Key Experimental Results

Main Results

Experiments are performed on PolitiFact and GossipCop from FakeNewsNet under a temporal split. Datasets comprise a 1:1 mixture of original articles and four stylistic rewrites generated by an LLM (sensational, objective, partisan, etc.), with a balanced 50:50 distribution of authentic and fake articles. Benchmarks include zero-shot LLMs (G1), fine-tuned SLMs (G2), and LLM-enhanced architectures (G3).

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

Ablation Study

To verify the contribution of each individual module under style-confusion attacks, systematic ablations were conducted on the test sets:

Model Variant PolitiFact (Acc) PolitiFact (F1) PolitiFact (Pre) PolitiFact (Rec) GossipCop (Acc) GossipCop (F1) GossipCop (Pre) GossipCop (Rec) Description
Ours (RATF) 0.8820 0.8823 0.8820 0.8819 0.7813 0.7816 0.7811 0.7807 Full proposed model
RATF w/o Rewritten Score 0.8487 0.8464 0.8482 0.8453 0.7526 0.7524 0.7532 0.7521 Removes \(s_{\text{rw}}\) condition
SC-CAF Only 0.8743 0.8754 0.8748 0.8759 0.7566 0.7561 0.7567 0.7558 Retains cross-attention, removes gating
Learnable Gate Only 0.8516 0.8507 0.8511 0.8502 0.7604 0.7615 0.7626 0.7612 Retains gating, removes cross-attention
Content-Only 0.8315 0.8308 0.8335 0.8315 0.7431 0.7429 0.7441 0.7431 Uses only \(A_{\text{con}}\) stream
Style-Only 0.7753 0.7744 0.7862 0.7753 0.7052 0.7044 0.7067 0.7052 Uses only \(A_{\text{sty}}\) stream

Key Findings

  • Rewritten Score Acts as an Essential Control Hub: Removing the rewritten score signal leads to substantial drops on both PolitiFact (F1: 0.8823 -> 0.8464, -3.59%) and GossipCop (F1: 0.7816 -> 0.7524, -2.92%), validating that unsupervised curvature estimation is essential for guiding adaptive fusion.
  • Extreme Fragility of Pure Stylistic Signals: The Style-Only model collapses to 0.7744 F1 on PolitiFact and 0.7044 on GossipCop, illustrating that stylistic markers cannot independently sustain veracity prediction under adversarial LLM manipulation.
  • No Penalty on Clean Benchmarks: Evaluated on unperturbed, clean test sets, RATF achieves 92.26% F1 on PolitiFact (outperforming Sheepdog's 88.39% and TED's 90.37%) and 76.51% F1 on GossipCop. This demonstrates that RATF eliminates the trade-off of style-invariance methods by retaining stylistic indicators whenever texts are authentic.
  • Strong Backbone Portability: Equipping different encoders with RATF yields consistent gains across RoBERTa (F1 from 76.34% to 87.85%), BERT (F1 from 75.36% to 86.31%), and DeBERTa (F1 from 77.02% to 86.72%).

Highlights & Insights

  • Continuous Calibration Over Binary Trade-Offs: By transforming the defense into a smooth probability-conditioned gating problem, RATF avoids the false dichotomy between full trust and blind discarding of style features.
  • Single-Pass Inference Architecture: Unlike multi-turn debate models (e.g., TED) or iterative rationale generators (e.g., ARG), RATF requires only a single offline LLM call to extract content and style streams, keeping online serving lightweight.
  • Empirical Disentanglement in Latent Space: t-SNE visualizations confirm that the tri-encoder constructs three non-overlapping, well-separated clusters for semantic, content, and style representations, validating the architectural hypothesis.

Limitations & Future Work

  • Reliance on External Foundation Models: RATF requires an LLM for content rewriting and a causal LM for curvature calculation, which increases pre-computation latency compared to lightweight single-encoder classifiers in streaming scenarios.
  • Limited Domain Diversity: Evaluations are confined to political claims (PolitiFact) and entertainment news (GossipCop); broader generalization across financial, medical, and multi-lingual news environments remains to be explored.
  • Lack of Multimodal and Retrieval Integration: Practical misinformation often pairs synthetic captions with authentic photos; future research should incorporate retrieval-augmented verification (RAG) and multimodal encoders.
  • vs Sheepdog (Wu et al., 2024): Sheepdog enforces style-invariance by training across multiple stylized reframings, completely discarding style cues and capping performance on clean news. RATF estimates style authenticity dynamically, preserving style utility on unperturbed articles.
  • vs ARG (Hu et al., 2024) / TED (Liu et al., 2025): ARG uses rationales that can be corrupted by deceptive rhetoric, while TED relies on multi-agent debate that is easily misled by formal prose. RATF anchors evidence retrieval in objective content queries, making it immune to rhetorical deception.

Rating

  • Novelty: ⭐⭐⭐⭐☆ Elegantly adapts conditional probability curvature from AI-text detection into a dynamic gating controller for misinformation defense.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive benchmarks spanning varying attack ratios, clean datasets, backbone adaptations, and component ablations.
  • Writing Quality: ⭐⭐⭐⭐⭐ Methodological progression is coherent and supported by transparent mathematical formulations and visual evidence.
  • Value: ⭐⭐⭐⭐☆ Highly practical and modular framework for defending automated fact-checking systems against generative AI manipulation.