[Paper Review] A Reputation Mechanism Is All You Need: Collaborative Fairness and Adversarial Robustness in Federated Learning

Xinyi Xu, Lingjuan Lyu|arXiv (Cornell University)|Nov 20, 2020

Privacy-Preserving Technologies in Data44 references39 citations

TL;DR

The paper introduces Robust and Fair Federated Learning (RFFL), a reputation-based framework that rewards participants according to their contributions and rejects adversaries, achieving collaborative fairness and robustness without extra validation data.

ABSTRACT

Federated learning (FL) is an emerging practical framework for effective and scalable machine learning among multiple participants, such as end users, organizations and companies. However, most existing FL or distributed learning frameworks have not well addressed two important issues together: collaborative fairness and adversarial robustness (e.g. free-riders and malicious participants). In conventional FL, all participants receive the global model (equal rewards), which might be unfair to the high-contributing participants. Furthermore, due to the lack of a safeguard mechanism, free-riders or malicious adversaries could game the system to access the global model for free or to sabotage it. In this paper, we propose a novel Robust and Fair Federated Learning (RFFL) framework to achieve collaborative fairness and adversarial robustness simultaneously via a reputation mechanism. RFFL maintains a reputation for each participant by examining their contributions via their uploaded gradients (using vector similarity) and thus identifies non-contributing or malicious participants to be removed. Our approach differentiates itself by not requiring any auxiliary/validation dataset. Extensive experiments on benchmark datasets show that RFFL can achieve high fairness and is very robust to different types of adversaries while achieving competitive predictive accuracy.

Motivation & Objective

• Motivate the need for collaborative fairness in federated learning where contributions vary across participants.
• Propose a reputation-based mechanism to quantify and reward contribution while discarding low-reputation participants.
• Address adversarial robustness to handle free-riders and malicious participants without auxiliary datasets.
• Demonstrate that RFFL achieves high fairness and robustness with competitive predictive accuracy on benchmark datasets.

Proposed method

• Define contributions v via standalone local test accuracy.
• Define rewards φ via final test accuracy; measure fairness with Pearson correlation ρp(v, φ).
• Introduce reputation r_i(t) for each participant; update via cosine similarity between uploaded gradient and aggregated gradient.
• Perform reputation-weighted gradient aggregation over reputable participants; normalize gradients to prevent explosion (γ).
• Control gradient download via quota_i proportional to r_i(t) to implement selective information exchange.
• Iteratively remove participants with reputation below threshold β to enforce robustness and fairness.

Experimental results

Research questions

• RQ1 Can a reputation mechanism align rewards with actual contributions in federated learning to achieve collaborative fairness?
• RQ2 Does reputation-weighted aggregation improve robustness against targeted/untargeted poisoning and free-riders without extra validation data?
• RQ3 What is the impact of reputation dynamics and thresholding on predictive accuracy and fairness under heterogeneous data settings?
• RQ4 How does RFFL compare to existing fairness and robustness baselines under various attack scenarios?

Key findings

• RFFL achieves higher fairness (as measured by Pearson correlation with standalone performance) than several baselines across multiple datasets and splits.
• Reputation-weighted aggregation improves predictive performance, especially under heterogeneous data splits (e.g., POW and CLA).
• RFFL is robust to untargeted poisoning and free-riders and can reduce the impact of targeted poisoning, though fully removing targeted attackers can be challenging under some configurations.
• Adversaries and free-riders often see their reputations drop below β and are removed from subsequent rounds, enhancing long-term robustness.
• CFFL can outperform RFFL in some settings but requires an auxiliary validation dataset; RFFL achieves competitive fairness without extra data.