Research on Defense Algorithm of Label Flipping Poisoning Attack Based on Dynamic Threshold Weighted KNN

As a typical data poisoning attack, label flipping poisoning attack destroys the performance of the model by tampering with the training data labels, which poses a serious threat to the machine learning system. To solve this problem, a dynamic threshold weighted KNN data disinfection algorithm is proposed. Based on the traditional KNN algorithm, the algorithm introduces the distance weighted voting and dynamic threshold mechanism. By calculating the similarity and category confidence between samples, it can effectively identify and filter the poisoned samples. Experiments verify its effectiveness in the confrontation environment. The experimental results show that the optimized KNN algorithm can more accurately identify the poisoning samples, improve the accuracy of the model, reduce the loss value of the model in the training process, and converge faster in different poisoning rate scenarios, showing more robust defense performance.