TL;DR
PrivFairFL resolves the inherent conflict between group fairness and privacy in Federated Learning (FL) by combining Secure Multiparty Computation (MPC) and Differential Privacy (DP). It enables training fair models in "cross-device" settings—where individual users hold sensitive data—without requiring them to disclose sensitive attributes like gender or ethnicity to a central authority.
The Challenge: Fairness vs. Privacy in Cross-Device FL
Machine learning models often inherit or amplify biases present in training data, leading to discriminatory outcomes in sensitive areas like hiring or credit approvals. "Group fairness" techniques aim to mitigate this, but they typically require access to sensitive attributes (e.g., race, gender) to calculate bias-correction terms or rebalance datasets.
This creates a paradox in Federated Learning. FL is designed to protect privacy by keeping data on-device. However, if the central aggregator cannot see sensitive attributes, it cannot easily measure or fix unfairness. Previous attempts to solve this either assumed a "cross-silo" setup (where large organizations hold data for many people) or sacrificed user privacy by leaking fairness metrics.
As shown in the comparison above, PrivFairFL uniquely targets the "cross-device" setup (dev), providing formal privacy guarantees (DP and MPC) for both individual data and aggregated statistics.
Proposed Framework: PrivFairFL-Pre and PrivFairFL-Post
The authors propose two main strategies that can be used independently or together, regardless of the specific FL training algorithm used:
- PrivFairFL-Pre (Pre-processing): This method assigns weights to training samples to balance the dataset before training begins. Clients send "secret shares" of their local counts (label + sensitive attribute) to a set of computing servers. These servers use Secure Multiparty Computation (MPC) to sum the counts without any single server seeing the raw numbers.
- PrivFairFL-Post (Post-processing): This method adjusts classification thresholds for different groups after the model is trained. The servers run an MPC protocol to construct ROC curves across the entire federation.
The Privacy "Secret Sauce"
To provide formal Differential Privacy (DP), the MPC servers cooperatively generate Laplacian noise and add it to the aggregated results before they are revealed. This effectively implements "Global DP"—which offers much better model utility than "Local DP" (where users add noise themselves)—but without needing a single trusted central party.
By adding noise inside the MPC protocol (as detailed in Protocol 1), the system achieves the high accuracy of global differential privacy while maintaining the decentralized trust of MPC.
Key Insights and Results
- Agnostic to Training: Because these are pre- and post-processing techniques, they are "plug-and-play" and can be combined with existing private training methods like DP-SGD.
- Efficiency in Scale: The pre-processing step (PrivFairFL-Pre) scales linearly with the number of clients, handling hundreds of thousands of clients in less than a second.
- Fairness Gains: In the ADS dataset (predicting user interest in ads), PrivFairFL-Pre significantly reduced the Equal Opportunity Difference ($\Delta EOP$) from 0.089 to 0.018.
- Utility vs. Privacy: Compared to baseline methods using Local DP, PrivFairFL achieved much higher accuracy. Local DP requires adding massive amounts of noise to every individual update, whereas PrivFairFL only adds noise once to the final aggregate.
Limitations and Future Work
While the pre-processing is efficient, the post-processing MPC protocols remain a bottleneck due to the high number of secure multiplications required to generate ROC curves (taking about 28 hours for 100,000 samples). Future work aims to develop in-processing bias mitigation techniques and optimize MPC protocols using "offline phases" to reduce runtime.
Summary updated