Explainable Disentangled Representation Learning for Generalizable Authorship Attribution in the Era of Generative AI

1. Problem Statement

The paper addresses two critical challenges in authorship analysis and AI-generated text detection:

• Content-Style Entanglement (Topic Confusion): Existing methods often learn spurious correlations between an author's identity and the topics they discuss (e.g., an author writing about "Detective Fiction" is misidentified as a specific author simply because of the topic, not their unique writing style). This leads to poor generalization across domains.
• Lack of Interpretability: Current neural approaches (even those using contrastive learning) treat style and content as a single embedding, offering no insight into which stylistic features drive attribution decisions.
• Generalization in the Age of LLMs: As Large Language Models (LLMs) generate increasingly fluent text, distinguishing human style from machine generation requires robust, topic-agnostic style representations that can generalize to few-shot or zero-shot scenarios.

2. Methodology: EAVAE

The authors propose EAVAE (Explainable Authorship Variational Autoencoder), a framework designed to explicitly disentangle authorial style from content through "separation-by-design." The architecture follows a two-stage training process:

A. Contrastive Pre-training

• Goal: Establish strong foundational authorial representations.
• Mechanism: A style encoder (based on LLMs) is pretrained using Supervised Contrastive Learning on a massive, diverse corpus (27.4M documents).
• Hard Negative Mining: To prevent the model from relying on surface-level lexical similarities, the authors use BM25 to retrieve negative samples (documents by different authors) that are lexically similar but stylistically distinct.
• Bidirectional Attention: The encoder utilizes bidirectional attention mechanisms to capture full contextual dependencies, improving representation quality.

B. Explainable Variational Autoencoder (VAE) Fine-tuning

This stage explicitly separates style and content using a specialized architecture:

1. Dual Encoders:
   • Style Encoder ($E_s$): Maps documents to a latent style space ($z_s$).
   • Content Encoder ($E_c$): Maps documents to a latent content space ($z_c$).
   • Independence Assumption: The model enforces $q(z_s, z_c|d) = q(z_s|d)q(z_c|d)$, ensuring the latent variables are statistically independent.
2. Shared Reconstructor: A single generator reconstructs the original text from the combined style and content latents ($z_s, z_c$), ensuring both components contain necessary information for reconstruction.
3. Explainable Discriminator:
   • Instead of a standard binary classifier, the discriminator is a generative model (using hybrid prompting).
   • Task: Given pairs of style or content representations, it classifies whether they originate from the same source (author or topic).
   • Explainability: Crucially, it generates natural language explanations justifying its decision (e.g., "These styles differ because of sentence structure X, while these contents are similar because of topic Y").
   • Hybrid Prompting: The generator uses fixed template prompts with learnable soft prompts, where placeholder tokens are dynamically replaced by the latent representations ($z_s, z_c$).

C. Training Objective

The total loss function combines:

• VAE Loss ($L_{vae}$): Reconstruction fidelity + KL divergence regularization for both style and content.
• Discriminator Loss ($L_{dis}$): A cross-entropy loss for classification accuracy plus the generation of coherent explanations.
• Total: $L_{EAVAE} = L_{vae} + \lambda_{dis} L_{dis}$.

3. Key Contributions

1. Architectural Disentanglement: A novel framework that explicitly separates style and content via dual encoders and a VAE structure, directly addressing the "content confounding" problem.
2. Explainable Adversarial Training: Introduction of a discriminator that not only enforces disentanglement but also provides natural language rationales for its decisions, enhancing model interpretability.
3. State-of-the-Art Performance: Demonstrated superior results in both Authorship Attribution (AA) and AI-generated text detection, proving that disentangled representations generalize better across diverse domains.

4. Experimental Results

The model was evaluated on multiple benchmarks:

• Authorship Attribution (AA):

  • Datasets: Amazon Reviews, PAN21, and the HRS corpus (5 heterogeneous domains).
  • Performance: EAVAE achieved State-of-the-Art (SOTA) results.
    • On Amazon Reviews: 97.0% MRR and 99.0% Recall@8 (surpassing LUAR by +3.6 MRR).
    • On HRS (Document-level): 47.3% MRR and 72.2% Recall@8. This was a massive improvement (+10.7 MRR) over the previous best (Man and Nguyen, 2024), highlighting the effectiveness of disentanglement in cross-domain scenarios.

• AI-Generated Text Detection:

  • Dataset: M4 Benchmark (ArXiv, PeerRead, WikiHow, Wikipedia).
  • Setting: Few-shot learning (Single-target and Multi-target detection).
  • Performance: EAVAE achieved competitive results without task-specific fine-tuning for detection.
    • Single-target: 65.7% pAUC@1, 93.5% pAUC@5, 98.5% pAUC@10.
    • Multi-target: 62.0% pAUC@1, 87.4% pAUC@5, 97.7% pAUC@10.
  • Significance: The model successfully detected AI text by relying on pure style representations, even when the content topics varied widely.

• Ablation Studies:

  • Removing the Disentanglement by Design (using a single encoder) caused the largest performance drop (MRR dropped from 47.3% to 44.5%), confirming that architectural separation is the most critical component.
  • Removing the Explainable Discriminator also significantly reduced performance, proving its role in enforcing independence.
  • The Hybrid Prompting mechanism outperformed simple MLP discriminators and soft-prompt-only variants.

5. Significance and Conclusion

• Robustness: EAVAE proves that explicitly separating style from content is essential for building authorship models that generalize across different topics and domains, a critical requirement in the era of generative AI.
• Interpretability: By generating natural language explanations, the model moves beyond "black box" predictions, allowing researchers to understand why a specific style attribution was made.
• Versatility: The learned style representations are transferable, showing strong performance in detecting AI-generated text without needing specific training on that task.
• Future Directions: The authors suggest extending the framework to multilingual settings, incorporating additional stylistic dimensions (sentiment, formality), and applying it to other modalities like code or speech.

In summary, EAVAE represents a significant leap forward in authorship analysis by solving the long-standing issue of topic confusion through principled architectural design and providing the necessary transparency for real-world forensic and security applications.