Researchers from Zhejiang University have developed a novel defense mechanism, Gradient Purification Defense (GPD), to protect decentralized federated learning systems from poisoning attacks. This new approach is groundbreaking because it not only removes harmful components but also retains beneficial information in malicious parties. This unique method secures the model and significantly enhance its accuracy, even beyond what is possible in an ideal, attack-free environment. The findings were published in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature, 15 September 2026.
Decentralized federated learning (DFL) allows multiple parties to train a shared AI model without exposing their private data. However, this makes it vulnerable to malicious participants who can send “poisoned” updates to sabotage the system. Current defenses typically reject all data from a suspicious parties, discarding potentially valuable knowledge that only that source possesses.
GPD overcomes this limitation through a divide-and-conquer strategy, which decouples the model gradient from the model weights, using “gradient detoxification” and “weight intoxication”. The GPD algorithm first designs mechanisms to eliminate the malicious influence in the gradient while retaining the malicious values in the model weights. Then, it uses the purified gradient to optimize the poisoned model weights. This optimization effectively retains the beneficial information embedded within the poisoned weights introduced by malicious parties., which aligns with the objective function and contributes positively to the final model performance.
This gradient purification defense turns a vulnerability into a potential advantage. Extensive experiments show GPD significantly outperforms state-of-the-art defenses, especially in scenarios with heterogeneous data distributions, making collaborative AI more robust and more accurate."
Frontiers of Computer Science
Experimental study
Not applicable
Gradient purification: defense against data poisoning attack in decentralized federated learning
15-Sep-2026