Balancing Privacy and Fairness

In today’s era of big data, safeguarding user data is crucial, especially when vast amounts of user information are collected and analyzed to provide services. Differential Privacy (DP) as a provable privacy concept, has gained significant attention and practical application across industry leaders, including giants like Google, Apple, Microsoft, and even governmental bodies like the U.S. government, where it’s been playing a critical role in safeguarding sensitive census data. DP achieves this safeguard by strategically injecting controlled noise into data or query results, making it nearly impossible to identify any individual’s unique contribution to the outcome. Fairness, on the other hand, is all about making sure that the data statistics or analysis treat everyone fairly and without bias. It tries to ensure that no one is treated differently or unfairly because of who they are. Fairness addresses the concern of discrimination and bias that can emerge in data and algorithmic decision-making processes.

While differential privacy and fairness may initially appear as distinct concepts, they share an interconnected relationship. Notably, DP often relies on the strategic introduction of controlled noise into the outputs of computations to protect individual privacy. However, this addition of noise can have unintended consequences for fairness. Researchers have demonstrated that the pursuit of differential privacy, such as through the use of DP-SGD (Differential Privacy Stochastic Gradient Descent), can have a disparate impact on model accuracy. Importantly, this impact tends to disproportionately affect under-represented groups, thereby influencing the fairness of the resulting models. This observation highlights the nuanced relationship between privacy and fairness, emphasizing the importance of addressing both considerations simultaneously in the design and deployment of data-driven systems to ensure a more equitable and privacy-protected future.

We have undertaken the exploration of the relationships between differential privacy and fairness across various machine learning models and privacy protection algorithms. Our focus is on discovering the factors of differential privacy may cause the fairness issue, and designing algorithms that can mitigate the impact of differential privacy on fairness while striving to find a harmonious solution that balances both factors. Our ultimate goal is to create a reliable data statistics and analysis solution that ensures both privacy and fairness, fostering trust and reliability.

Details of the research themes include:

• Intersections between DP and fairness: Investigate the interplay between differential privacy and fairness across a variety of models and application scenarios. Analyze the factors influencing fairness within this context and delve into how fairness considerations may inadvertently impact either privacy protection or model accuracy in a reverse manner.
• Privacy-Fairness Trade-off: Investigate the trade-off between achieving high levels of privacy through DP and maintaining fairness in algorithmic outcomes. Explore how different levels of privacy protection impact the fairness of models and decisions and seek to strike an optimal balance.
• Fairness-aware DP Techniques: Develop and evaluate privacy-preserving techniques that explicitly consider fairness constraints. Modify and extend the DP algorithms to ensure that they do not disproportionately affect certain demographic groups or contribute to biased outcomes.