1 results about "Rank analysis" patented technology

The application relates to the field of artificial intelligence and discloses a federal learning gradient defense method based on label repetition rate. Through rank analysis of a relationship equation of gradient leakage data, it is found that if the same label exists in batch samples and the obtained label prediction probability is similar, the effect of the gradient leakage attack can be effectively reduced, that is, it is difficult to reconstruct the input sample from the gradient. Therefore, in the data set loading stage, the samples of the data set are placed in the corresponding label bucket according to the label category, then the samples are placed in the corresponding confidence layer according to the confidence in the label bucket, finally, the sample combination is formed according to the setting of the label repetition number, and the sample combination is shuffled to form a new data set. The federal learning gradient defense method based on the label repetition rate proves that the label repetition rate can defend against the gradient leakage attack in theory through rank analysis of the relationship equation of the gradient leakage data, and the method is verified in practical application.