Differential Privacy, Linguistic Fairness, and Training Data Influence: Impossibility and Possibility Theorems for Multilingual Language Models

Phillip Rust, Anders Søgaard

As multilingual language models like mBERT, XLM-R, and BLOOM become the foundational infrastructure for global AI applications, the demand for them to be "trustworthy" has never been higher. But what does trustworthiness actually entail? In the context of multilingual AI, it generally rests on three pillars: Differential Privacy (DP) to protect individual training examples, Linguistic Fairness to ensure equitable performance across different languages, and Transparency to explain predictions by tracing them back to specific training data.

This paper investigates a critical question: Can a single multilingual model satisfy all three requirements simultaneously? By combining theoretical proofs with extensive empirical experiments, the authors reveal the inherent tensions—and surprising synergies—between privacy, fairness, and transparency.

The Core Conflict: Privacy vs. Transparency

One of the paper's most significant contributions is its formal Impossibility Theorem. The authors prove that Differential Privacy and instance-based transparency (specifically, the sparsity of training data influence) are fundamentally incompatible.

If a model is truly private, it must guarantee that no single training data point has a disproportionate impact on the model's final parameters or outputs. Consequently, "instance-based" explanations—where a system justifies a prediction by pointing to a highly influential training example—are mathematically weakened by privacy guarantees. As a model becomes more $\epsilon$-differentially private, it inherently becomes less instance-interpretable. The more "multilingual" and compressed the representations, the harder it is to isolate a single influential training example.

The Surprising Synergy: Privacy and Fairness

While privacy and transparency clash, the authors established a Possibility Theorem demonstrating that privacy and linguistic fairness are theoretically compatible through the mechanism of multilingual compression.

"Perfect compression" occurs when a model represents the same semantic concept identically, regardless of the input language. The authors show that this ideal state naturally leads to $k$-anonymity, which in turn facilitates differential privacy. In short, forcing a model to merge languages into a shared representation space can actually help protect the privacy of the underlying data.

Experimental Insights: The "U-Shape" Trade-off

To test these theories, the researchers fine-tuned XLM-R and mBERT using DP-AdamW (a private optimizer) across various privacy budgets ($\epsilon \in \{1, 3, 8, 15, 30, \infty\}$). They evaluated the models on Part-of-Speech (POS) tagging (PUD treebank) and Natural Language Inference (XNLI), using parallel corpora from TED 2020, WikiMatrix, and Tatoeba across 15+ typologically diverse languages.

To measure how well the models compressed languages, they utilized proxy metrics like Sentence Retrieval, Centered Kernel Alignment (CKA), IsoScore, and Representational Similarity Analysis (RSA). To measure data influence, they used TracInCP (tracing gradient changes) and introduced a new metric called InfU (Influence Uniformity) to quantify how spread out training data influence is.

Key empirical findings include:

  • The Compression Valley: Multilingual compression follows a non-linear, "U-shaped" path relative to privacy. High privacy ($\epsilon=1$) and no privacy ($\epsilon=\infty$) both exhibited high compression, but for entirely different reasons. At $\epsilon=1$, the heavy noise added for privacy forces the model to ignore language-specific nuances, leading to high compression but poor task performance. Interestingly, at medium privacy levels (e.g., $\epsilon=8$), models fell into a "compression valley"—they retained enough signal to be language-specific but had too much noise to allow clean cross-lingual mapping.
  • Performance vs. Privacy: As expected, tightening privacy degrades accuracy. On the XNLI benchmark, the non-private model achieved roughly 70% test accuracy, whereas the high-privacy model ($\epsilon=1$) plummeted to about 36%, which is near random guessing.
  • DP as a Regularizer: Despite the drop in overall accuracy, private models exhibited a much smaller gap between validation and test performance compared to non-private models, indicating that Differential Privacy acts as a strong form of regularization.

Limitations and the Path Forward

While these findings provide a crucial theoretical foundation, the authors note several limitations. The study focused strictly on classification tasks (POS, NLI), and generative tasks like translation or summarization may exhibit entirely different trade-off dynamics. Furthermore, "perfect compression" remains a theoretical ideal; current metrics like CKA are merely proxies for this state.

Ultimately, the results highlight a pressing need for novel optimization techniques. Since DP-tuning is notoriously sensitive to hyper-parameters, future research must focus on methods that can jointly target performance, privacy, and fairness without forcing developers to sacrifice one trustworthy pillar for another.

Summary updated

Abstract

Language models such as mBERT, XLM-R, and BLOOM aim to achieve multilingual generalization or compression to facilitate transfer to a large number of (potentially unseen) languages. However, these ...
BibTeX
@inproceedings{rust2023differential,
  author = {Phillip Rust and Anders Søgaard},
  title = {{Differential Privacy, Linguistic Fairness, and Training Data Influence: Impossibility and Possibility Theorems for Multilingual Language Models}},
  booktitle = {International Conference on Machine Learning},
  year = {2023},
  url = {https://openreview.net/forum?id=SP6w4sVCyp},
}
Figure