DABS: Data-Agnostic Backdoor attack at the Server in Federated Learning

• DOI: 10.48550/arxiv.2305.01267
• Publication Date: 2023-01-01
• Authors: Wenqiang Sun; Sen Li; Yuchang Sun; Jun Zhang
• Keywords: Backdoor; Attack model

Federated learning (FL) attempts to train a global model by aggregating local models from distributed devices under the coordination of central server. However, existence large number heterogeneous makes FL vulnerable various attacks, especially stealthy backdoor attack. Backdoor attack aims trick neural network misclassify data target label injecting specific triggers while keeping correct predictions on original training data. Existing works focus client-side attacks which try poison modifying datasets. In this work, we propose new for FL, namely Data-Agnostic at Server (DABS), where server directly modifies an system. Extensive simulation results show that scheme achieves higher success rate compared with baseline methods maintaining normal accuracy clean