Temporally Sparse Attack against Large Language Models in Time Series Forecasting

Fuqiang Liu, Sicong Jiang

Keywords: large language models, adversarial attack, AI safety

TL;DR: This study introduces a novel temporally sparse adversarial attacks, showing that manipulate a small part of input time series can significantly degrade LLM-based forecasting performance.

Abstract: Large Language Models (LLMs) have recently demonstrated strong potential in zero-shot time series forecasting by leveraging their ability to capture complex temporal patterns through the next-token prediction mechanism. However, recent studies indicate that LLM-based forecasters are highly sensitive to small input perturbations. Existing attack methods, though, typically require modifying the entire time series, which is impractical in real-world scenarios. To address this limitation, we propose a Temporally Sparse Attack (TSA) against LLM-based time series forecasting. We formulate the attack as a Cardinality-Constrained Optimization Problem (CCOP) and introduce a Subspace Pursuit (SP)--based algorithm that restricts perturbations to a limited subset of time steps, enabling efficient and effective attacks. Extensive experiments on state-of-the-art LLM-based forecasters, including LLMTime (GPT-3.5, GPT-4, LLaMa, and Mistral), TimeGPT, and TimeLLM, across six diverse datasets, demonstrate that perturbing as little as 10% of the input can substantially degrade forecasting accuracy. These results highlight a critical vulnerability of current LLM-based forecasters to low-dimensional adversarial attacks.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 23681