A Secure federated learning framework based on autoencoder and Long Short-Term Memory with generalized robust loss function for detection and prevention of data poisoning attacks

• To suggest a federated learning-based framework for detecting poisonous data. • To perform data cleaning for handling the missing values to improve data quality. • To handle data privacy issues via federated learning by training the deep network. • To suggest an AE-LSTM with robust loss function for detecting the data poisonous attack. In this research, a federated learning-based poisoning attack recognition and prevention framework has been developed. Initially, the required data to perform data poison attack detection is gathered from the standard dataset. Next, data cleaning takes place on the collected data to remove the corrupted data. The Federated Learning Network (FedNet) is used for the detection of poison attacks. Using FedNet, the Autoencoder and Long Short Term Memory (AE-LSTM) with novel loss functions named generalized robust loss functions are designed for detecting the poison attack in the cleaned data without any loss of information. The recommended federated learning-based AE-LSTM with novel generalized robust loss functions detects and prevents the poison attack to protect the sensitive information in the network. Finally, the validation is conducted to verify the performance of the developed poison attack detection framework.