Practical Implementation of Federated Learning for Detecting Backdoor Attacks in a Next-word Prediction Model.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) allows collaborative model training across decentralized devices, preserving data privacy.
• FL systems are vulnerable to backdoor attacks, where malicious participants inject biased data to manipulate model outputs.
• Such attacks can significantly skew model performance, particularly on sensitive topics like elections.

Purpose of the Study:

• To develop and evaluate a novel mechanism for detecting backdoor attacks in federated learning.
• To enhance the robustness and reliability of next-word prediction models trained via federated learning.
• To mitigate the impact of malicious data manipulation on model outputs.

Main Methods:

• Development of a federated learning framework for next-word prediction.
• Implementation of a detection mechanism to identify and exclude devices with anomalous datasets.
• Experimental validation using a presidential election scenario to quantify attack impact and detection effectiveness.

Main Results:

• A positive correlation was observed between the proportion of compromised devices and the degree of model bias.
• The proposed detection mechanism effectively reduced the impact of backdoor attacks, especially with a low percentage of malicious devices.
• The system demonstrated improved robustness against targeted data poisoning.

Conclusions:

• The developed detection mechanism significantly enhances the security and reliability of federated learning systems.
• This research contributes to building more trustworthy AI models in decentralized environments.
• The findings underscore the importance of robust defense strategies against adversarial manipulation in federated learning.