Using Unsupervised Methods to Find Attack Points for Time Series Prediction Models
Qi Zou · Artificial intelligence technology research. · 2024
Time series prediction models are extensively utilized across different industries in everyday life, and the security of industry data is linked to the susceptibility of these prediction models to adversarial attacks. This paper focuses on the method of fi nding vulnerable positions for adversarial attacks in time series prediction problems. Currently, most related studies use supervised methods to fi nd attack points, but they require labeled training data, which may be diffi cult to obtain or costly in some cases, and they may not adapt well to new attacks because the model only learns the features of known attacks during training. Therefore, in this paper, We defi ne the common vulnerable position identifi ed by the brute force method as the real common vulnerable position. Next, we introduce the application of unsupervised methods (such as Kmeans) to identify common vulnerable positions. We compare the positions identifi ed by these two methods and validate them using an LSTM time series prediction model.