Reconstruction-Error Based Data-Efficient Backdoor Attacks for Multivariate Time Series Forecasting

doi:10.1109/lsp.2026.3736349

Reconstruction-Error Based Data-Efficient Backdoor Attacks for Multivariate Time Series Forecasting

2026-09-21

0

PRE

AI

DOI：10.1109/lsp.2026.3736349

Original

Original request for help

Save

Abstract

En 中文

Deep neural networks (DNNs) have been widely applied in multivariate time series (MTS) forecasting. The backdoor attack, which can manipulate model behavior and mislead downstream decisions by poisoning part of the training data, remains a challenging problem for model security. Recent studies reveal that the efficiency of poison sample selection is critical in backdoor attacks, as different samples induce different parameter updates during DNN training, ultimately influencing attack effectiveness. Existing selection methods typically define a scoring metric to evaluate different training samples and then select samples according to their scores. However, for MTS data, the poison variables within each sample remain less explored. In this letter, we show that poisoning variables and temporal segments with different reconstruction-error (RE) yields different attack effectiveness. Guided by this observation, we propose DEBA-MTS, a straightforward yet efficient sample selection strategy. DEBA-MTS obtain RE scores along the variable and temporal segments perspective and select samples by RE difference. Experiments on real-world datasets and mainstream forecasting models demonstrate the effectiveness of DEBA-MTS.

Keywords:

Multivariate time series forecasting

backdoor attack

data-efficient

reconstruction error

data security

Reconstruction-Error Based Data-Efficient Backdoor Attacks for Multivariate Time Series Forecasting

Abstract

Journal

Organization

Cited Papers