TL;DR
This paper explores the challenging intersection of privacy-preserving techniques for sensitive text data. They investigate combining Federated Learning (FL) with Differential Privacy (DP) for Named Entity Recognition (NER), a crucial step in anonymizing text. While their work is the first to explore this specific combination for sequence tagging, they find that the synergy of these privacy measures can lead to unstable models, highlighting the need for careful tuning and further research.
What Problem Does This Paper Address?
In today's data-driven world, vast amounts of text data are generated, offering immense potential for machine learning applications. However, for sensitive domains like healthcare and law, this text often contains private information that must be anonymized to protect individuals' privacy. Sequence tagging tasks, such as Named Entity Recognition (NER), are key to this anonymization process by identifying and classifying private information spans.
The challenge lies in training effective NER models. This requires a significant amount of labeled data, which, in privacy-sensitive domains, is often distributed across different data silos and cannot be directly shared due to legal regulations like HIPAA and GDPR. Centralizing this data for traditional machine learning training poses a significant privacy risk.
Federated Learning (FL) offers a solution by allowing models to be trained on decentralized data without raw data ever leaving the client devices. However, FL itself is not immune to privacy attacks. Differential Privacy (DP) provides a strong, mathematically guaranteed privacy guarantee. This paper investigates the combined application of FL and DP for NER, aiming to achieve robust privacy while maintaining model performance.
Methods
The authors propose a framework that integrates two levels of privacy protection for sequence tagging, specifically NER:
- Federated Learning (FL): They adapt the widely used FederatedAveraging algorithm. In this setup, a global model is shared from a central server to multiple client sites. Each client trains the model on its local, private data. Instead of sharing the data, clients send only their model parameter updates back to the server, which then aggregates these updates to improve the global model. This process is repeated for several communication rounds.
- Differential Privacy (DP): To further enhance privacy, they incorporate DP into the client-side training process. Specifically, they use the Differentially Private Stochastic Gradient Descent (DP-SGD) optimizer. This involves:
- Computing gradients for a micro-batch of training data.
- Clipping these gradients to limit their influence.
- Adding carefully calibrated random noise to the clipped gradients.
- Using these noisy gradients to update the model parameters.
They experiment with two NER model architectures: a BiLSTM with a TimeDistributed Dense Layer (TDDL) as the final layer, and a BiLSTM with a Conditional Random Field (CRF) as the final layer.
Datasets
The experiments were conducted using the CoNLL 2003 English NER dataset. This dataset includes:
- Training set: 14,987 tokens
- Validation set: 3,466 tokens
- Test set: 3,684 tokens
The tokens are tagged with one of four entity types (Person, Location, Organization, Miscellaneous) or labeled as 'Other' (O), following the BIO2 annotation standard, resulting in 9 possible labels per token.
Key Insights and Ideas
- The Synergy of FL and DP is Complex: The paper's core finding is that while both FL and DP are valuable for privacy, their combination for sequence tagging tasks like NER can lead to unstable models and performance degradation. This is the first study to investigate this specific combination for this task.
- CRF Architecture Struggles with DP: The BiLSTM-CRF model architecture showed a significant performance drop and fluctuation when DP was applied, suggesting that CRF-based models might be more sensitive to the noise introduced by DP-SGD.
- FL Performance Depends on Client Data Size: In a standard FL setup (without DP), the performance of the aggregated model decreased as the number of clients increased. This is intuitive, as each client has less data to train on.
- Incremental FL Shows Promise: An incremental FL approach, where clients train on a small percentage of their local data and communicate updates frequently, showed promising results and maintained reasonable performance, making it suitable for scenarios where large amounts of data are not available at once on clients.
- DP's Impact on Incremental FL: When DP was introduced into the incremental FL setup, performance degraded, especially as the percentage of training data used per client increased.
Key Results
- Baseline NER Performance: The BiLSTM-TDDL model achieved an F-measure comparable to state-of-the-art results (0.916 vs. 0.9309). The BiLSTM-CRF model performed slightly lower (0.864).
- DP without FL: Applying DP-SGD to the BiLSTM-TDDL model showed stable performance with varying noise multipliers, indicating its robustness. However, the BiLSTM-CRF model's performance deteriorated significantly with DP.
- FL without DP: Increasing the number of clients in the FL setup led to a decrease in F-measure (e.g., from 0.874 with 2 clients to 0.766 with 20 clients). The incremental FL approach achieved an F-measure of 0.846 with 100% data.
- FL with DP: When DP was applied in the FL setup, the F-measure decreased by approximately 1-6% compared to the non-private FL performance. In the incremental FL setup with DP, performance dropped significantly after using 60% of the training data.
Limitations and Future Work
- Model Instability: The primary limitation identified is the instability and performance degradation observed when combining FL and DP, particularly for the CRF architecture.
- Root Cause Analysis: The authors acknowledge the need to investigate the root cause of the performance drop in CRF-based models with DP and to find ways to mitigate this issue.
- Hyperparameter Tuning: Further research is needed to explore optimal hyperparameter settings for both FL and DP to maximize performance.
- Dataset Generalization: The study was conducted on a single NER dataset. Future work should involve testing on other NER datasets, including low-resource ones, to assess the generalizability of the findings.
- Broader Goal: The ultimate goal is to build a robust differentially private federated framework for sequence tagging tasks that minimizes performance compromise.
Summary updated