Abstract
Existing studies have demonstrated that adversarial examples can be directly attributed to the presence of non-robust features, which are highly predictive, but can be easily manipulated by adversaries to fool NLP models. In this study, we explore the feasibility of capturing task-specific robust features, while eliminating the non-robust ones by using the information bottleneck theory. Through extensive experiments, we show that the models trained with our information bottleneck-based method are able to achieve a significant improvement in robust accuracy, exceeding performances of all the previously reported defense methods while suffering almost no performance drop in clean accuracy on SST-2, AGNEWS and IMDB datasets.
Introduction
Deep learning models in Natural Language Processing (NLP) are notoriously fragile. Small, semantics-preserving changes—such as replacing a word with its synonym (TextFooler) or swapping a few characters (TextBugger)—can completely flip a model's prediction. This vulnerability stems from "non-robust features": patterns in the data that are highly predictive for a specific dataset but are easily manipulated by an attacker. For example, a sentiment classifier might learn to rely on specific stylistic quirks rather than the actual sentiment of the words. The challenge is to build a "filter" that allows robust, task-relevant features to pass through while blocking the noisy, non-robust ones that adversaries exploit.
Methodology: The Information Bottleneck Approach
To address this, we propose a novel implementation of the Information Bottleneck (IB) principle. The IB principle seeks to find a representation $T$ that is as compressed as possible relative to the input $X$ (minimizing mutual information $I(X; T)$) while remaining as informative as possible about the target label $Y$ (maximizing mutual information $I(Y; T)$).
The IB Layer
Unlike previous attempts (such as InfoBERT) that placed a bottleneck at the initial embedding layer, our method inserts an IB layer between the final BERT output and the classifier. This placement is strategic; because BERT's higher layers contain more task-specific, contextualized information, the "filtering" process becomes more precise.
Variational Inference and Control
Since calculating mutual information is computationally difficult, we utilize a variational approximation. We minimize a loss function that combines standard Cross-Entropy (for prediction) with a Kullback-Leibler (KL) divergence term (for compression). A hyperparameter $\beta$ controls the "neck" of the bottleneck: a higher $\beta$ narrows the bottleneck, forcing more compression and filtering out more features. This method is "plug-and-play," meaning it can be combined with existing defenses like FreeLB (adversarial training) to further boost performance.
Experimental Setup
We evaluated the approach on three standard text classification benchmarks:
- SST-2: A binary sentiment analysis dataset of movie reviews.
- IMDB: A large-scale movie review sentiment dataset.
- AGNEWS: A four-class topic classification dataset (World, Sports, Business, Sci/Tech).
Key Results and Insights
Massive Robustness Gains
The IB-based model outperformed several state-of-the-art (SOTA) defense methods across all datasets and attack types. On the IMDB dataset under the TextFooler attack, a standard BERT-base model's accuracy plummeted to 0.8%. Our proposed IB model maintained 52.2% accuracy, which jumped to 64.3% when combined with FreeLB. Against character-level attacks like Deepwordbug, the model achieved 71.3% robust accuracy on IMDB, compared to just 24.3% for the baseline BERT model.
Preservation of Clean Accuracy
Historically, making a model robust often meant sacrificing "clean" accuracy (performance on un-attacked data). Our findings show that by filtering out irrelevant features, we can maintain—or even slightly improve—clean accuracy while drastically increasing robustness. In almost all tests, the accuracy on non-adversarial data remained within 0.1-0.5% of the baseline.
Feature Significance
By calculating "significance scores" for words, we demonstrated that the IB-enhanced model successfully ignores "stop words" and irrelevant modifiers (like "perhaps" or "the") and focuses its attention on high-impact sentiment words (like "joyless" or "slipshod"). Visualizations using t-SNE further confirmed that as the bottleneck narrows, the boundaries between classes become clearer and more distinct.
Conclusion and Limitations
This study demonstrates that the Information Bottleneck is a powerful tool for enhancing NLP robustness without the typical performance trade-offs. However, the model relies on finding the "sweet spot" for the hidden dimension size and the $\beta$ coefficient. If the bottleneck is too narrow, the model loses task-essential information. Future work will explore how this IB implementation performs on more complex NLP tasks like Natural Language Inference (NLI) or Question Answering (QA).
Summary updated