Summary of Dp-dylora: Fine-tuning Transformer-based Models On-device Under Differentially Private Federated Learning Using Dynamic Low-rank Adaptation, by Jie Xu et al.
DP-DyLoRA: Fine-Tuning Transformer-Based Models On-Device under Differentially Private Federated Learning using Dynamic Low-Rank Adaptation
by Jie Xu, Karthikeyan Saravanan, Rogier van Dalen, Haaris Mehmood, David Tuckey, Mete Ozay
First submitted to arxiv on: 10 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Cryptography and Security (cs.CR); Distributed, Parallel, and Cluster Computing (cs.DC)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Federated learning enables collaborative model training without sharing local data. However, this approach can still leak sensitive information. Differential privacy (DP) addresses this issue by providing formal guarantees against leakage through randomness. This paper empirically evaluates the practicality of fine-tuning large transformer-based models with DP in federated learning systems. Experiments on various domains, including speech recognition, computer vision, and natural language understanding, demonstrate that full fine-tuning under DP leads to significant performance degradation. Reducing dimensionality through parameter-efficient fine-tuning (PEFT) alleviates this issue. The proposed method, DP-Low-Rank Adaptation (DP-LoRA), consistently outperforms existing methods. Additionally, the paper introduces an adaptation method called DP-DyLoRA that can be combined with differential privacy to reduce accuracy degradation and word error rate. The results show that with 1 million clients and a stringent privacy budget of , the accuracy degradation and WER increase are reduced to less than 2% and 7%, respectively. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making sure that when people work together to train a machine learning model, they don’t accidentally share sensitive information. They do this by adding randomness to the data they contribute. The problem is that this makes it hard to train big models. The researchers tested different ways of training these big models and found that one way, called parameter-efficient fine-tuning (PEFT), works better than others. They also introduced a new method that combines well with differential privacy to reduce errors. Overall, the paper shows that it’s possible to keep sensitive information private while still getting good results from machine learning models. |
Keywords
» Artificial intelligence » Federated learning » Fine tuning » Language understanding » Lora » Low rank adaptation » Machine learning » Parameter efficient » Transformer
