Adapt-LFA: Adaptive Gradient-Guided Label Flipping Attack Against Federated Learning-based Intrusion Detection in IoT

Adopting Federated Learning (FL) in Intrusion Detection Systems (IDS) for IoT enhances privacy by decentralizing model training. However, FL is still vulnerable to adversarial threats, particularly Label-Flipping Attacks (LFA) that manipulate local training data to degrade model performance. This paper introduces the Adaptive Gradient-Guided Label Flipping Attack (Adapt-LFA), which strategically flips training labels using gradient-based optimization to maximize classification errors while minimizing detection risks. Evaluations on the CSE-CICIDS2018 and CICIoV2024 datasets show that Adapt-LFA reduces accuracy by $10 \%$ in Recurrent Neural Networks (RNN) and 13% in Convolutional Neural Networks (CNN), outperforming baseline LFA in disrupting FL-based IDS. These results highlight the effectiveness of the attack in degrading IDS performance within FL-based IoT environments.