Summary of De-amplifying Bias From Differential Privacy in Language Model Fine-tuning, by Sanjari Srivastava et al.
De-amplifying Bias from Differential Privacy in Language Model Fine-tuning
by Sanjari Srivastava, Piotr Mardziel, Zhikhun Zhang, Archana Ahlawat, Anupam Datta, John C Mitchell
First submitted to arxiv on: 7 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Cryptography and Security (cs.CR); Computers and Society (cs.CY); Methodology (stat.ME)
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 This research paper explores the tension between fairness and privacy in machine learning. The authors examine how differential privacy (DP) mechanisms affect model bias when fine-tuning large language models. Surprisingly, they find that DP amplifies gender, racial, and religious bias, leading to more biased models than those without DP. The root cause is a disparity in gradient convergence across sub-groups. However, the authors demonstrate that Counterfactual Data Augmentation (CDA) can mitigate this bias amplification, allowing for fair and private model fine-tuning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Machine learning wants to be kind and protect your secrets! This paper talks about two important ideas: fairness and privacy. Fairness means making sure a computer program doesn’t have bad opinions about people because of their race, gender, or religion. Privacy is when computers keep what they learn from you private, so nobody else can use it against you. The problem is that these two ideas are hard to balance. The researchers found out that using special techniques called differential privacy makes things worse – it makes the computer program even more biased! But they also discovered a way to fix this by using another technique called Counterfactual Data Augmentation. |
Keywords
* Artificial intelligence * Data augmentation * Fine tuning * Machine learning
