[Paper Review] Style Obfuscation by Invariance

Chris Emmery, Enrique Manjavacas+1 more|arXiv (Cornell University)|May 18, 2018

Authorship Attribution and Profiling46 references3 citations

TL;DR

This paper proposes obfuscation-by-invariance, a novel approach to style obfuscation that trains sequence models to be invariant to specific stylistic attributes (e.g., age, gender) rather than transferring style. Using a Gradient Reversal Layer (GRL) in an autoencoder or sequence-to-sequence framework, the method achieves chance-level performance for style classifiers while preserving semantic content better than transfer-based methods, as validated by human evaluation showing a trade-off between obfuscation strength and output quality.

ABSTRACT

The task of obfuscating writing style using sequence models has previously been investigated under the framework of obfuscation-by-transfer, where the input text is explicitly rewritten in another style. These approaches also often lead to major alterations to the semantic content of the input. In this work, we propose obfuscation-by-invariance, and investigate to what extent models trained to be explicitly style-invariant preserve semantics. We evaluate our architectures on parallel and non-parallel corpora, and compare automatic and human evaluations on the obfuscated sentences. Our experiments show that style classifier performance can be reduced to chance level, whilst the automatic evaluation of the output is seemingly equal to models applying style-transfer. However, based on human evaluation we demonstrate a trade-off between the level of obfuscation and the observed quality of the output in terms of meaning preservation and grammaticality.

Motivation & Objective

• To address the semantic degradation common in style obfuscation via transfer, where models alter meaning to fool classifiers.
• To investigate whether training models to be invariant to stylistic attributes improves meaning preservation during obfuscation.
• To evaluate the effectiveness of Gradient Reversal Layers (GRL) in learning style-invariant representations for text.
• To compare performance of parallel (sequence-to-sequence) and non-parallel (autoencoder) settings in style obfuscation.
• To assess the trade-off between obfuscation strength and output quality using human evaluation.

Proposed method

• Uses a neural encoder-decoder architecture with a Gradient Reversal Layer (GRL) inserted between the sentence encoder and the style classifier to enforce style invariance.
• Trains the model to minimize style classifier accuracy, pushing it toward chance level, while preserving semantic content.
• Employs a conditioned decoder in both parallel and non-parallel settings to generate output based on the style-invariant latent representation.
• Applies the model to the English Bible corpus in both parallel (with paired style translations) and non-parallel (autoencoder) setups.
• Uses automatic metrics (BLEU, METEOR, embedding similarity) and human evaluation to compare outputs.
• Evaluates the impact of the GRL and conditional decoding on both adversary performance and semantic fidelity.

Experimental results

Research questions

• RQ1 Can a model trained to be invariant to stylistic attributes achieve effective style obfuscation without altering semantic content?
• RQ2 How does the performance of style-invariant models compare to style-transfer models in terms of meaning preservation and grammaticality?
• RQ3 To what extent does the use of a Gradient Reversal Layer reduce style classifier accuracy while maintaining semantic coherence?
• RQ4 Does the autoencoder setting with GRL produce more semantically stable rewrites than the sequence-to-sequence model with parallel data?
• RQ5 What is the trade-off between the level of obfuscation and the quality of the generated output as judged by humans?

Key findings

• The style classifier's performance was reduced to chance level, indicating successful style invariance.
• Automatic evaluation metrics (BLEU, METEOR, embedding similarity) showed outputs from the style-invariant model were comparable to those from style-transfer models.
• Human evaluation revealed a clear trade-off: higher obfuscation strength correlated with lower perceived grammaticality and meaning preservation.
• The autoencoder with GRL outperformed the sequence-to-sequence model with transfer in human evaluation, despite the latter having access to parallel data.
• Qualitative analysis showed that the autoencoder model produced semantically correct local changes and partial rewrites, suggesting potential for style-neutral rewrites.
• Evidence of style-neutral rewrites was difficult to find in the Bible corpus, but the autoencoder approach showed promise for non-parallel settings.