Securing Federated Learning

Federated Learning (FL) facilitates decentralized model training across various clients while safeguarding data privacy, but it poses significant security risks, such as backdoor attacks, poisoning attacks, Man-in-the-Middle (MITM) attacks, and label inference attacks. This research introduces a security framework with four core components, each tailored to mitigate specific threats in FL environments.

HydraGuard is a hybrid defense against backdoor attacks, combining Trigger Inversion and Simple Tuning methods. Trigger Inversion identifies potential backdoor triggers by reconstructing patterns from model gradients, while Simple Tuning adjusts specific layers of the model to neutralize these threats. This dual approach reduces backdoor attack success rates while maintaining model accuracy.

SECUNID addresses data and model poisoning attacks by filtering out malicious updates before they are integrated into the global model. It utilizes LayerCAM, a visual interpretability tool, and an Autoencoder to detect and exclude anomalies, ensuring only legitimate updates contribute to the model.

CodeNexa prevents MITM attacks by implementing dynamic metric verification, which encrypts and validates key performance metrics during FL, safeguarding the system from unauthorized alterations during communication.

SHIELD protects Vertical Federated Learning (VFL) systems from label inference attacks using confusional autoencoders, noise addition, and a simplified Secure Multi-Party Computation (SMC) framework. These techniques obscure data representations, reducing the risk of sensitive label inference.