EAB-FL: Exacerbating Algorithmic Bias through Model Poisoning Attacks in Federated Learning

Federated Learning has a hidden vulnerability to fairness attacks: While federated learning is praised for protecting privacy by keeping data decentralized, this paper reveals it can be manipulated by malicious participants to deliberately worsen bias against specific demographic groups—a threat that wasn't previously well understood.

Current fairness safeguards are insufficient: The research demonstrates that existing fairness optimization algorithms designed to protect against bias in federated learning can be circumvented by adversarial attacks, meaning current protections may give a false sense of security.

Attackers can hide their malicious intent: The EAB-FL attack is particularly dangerous because it maintains good overall model performance while simultaneously degrading fairness for certain groups, making the bias harder to detect than attacks that simply reduce accuracy.

The problem affects real-world collaborative AI systems: Since federated learning is used across multiple organizations and devices to train shared models, this vulnerability could impact AI systems used in hiring, lending, healthcare, and other high-stakes domains where demographic bias is already a concern.

Urgent need for stronger defenses: The paper's finding that attacks succeed "even with state-of-the-art fairness optimization algorithms" suggests that organizations deploying federated learning need to develop new security measures specifically designed to protect fairness, not just accuracy and privacy.