Multivariate time series (MTS) regression tasks are common in many real-world\ndata mining applications including finance, cybersecurity, energy, healthcare,\nprognostics, and many others. Due to the tremendous success of deep learning\n(DL) algorithms in various domains including image recognition and computer\nvision, researchers started adopting these techniques for solving MTS data\nmining problems, many of which are targeted for safety-critical and\ncost-critical applications. Unfortunately, DL algorithms are known for their\nsusceptibility to adversarial examples which also makes the DL regression\nmodels for MTS forecasting also vulnerable to those attacks. To the best of our\nknowledge, no previous work has explored the vulnerability of DL MTS regression\nmodels to adversarial time series examples, which is an important step,\nspecifically when the forecasting from such models is used in safety-critical\nand cost-critical applications. In this work, we leverage existing adversarial\nattack generation techniques from the image classification domain and craft\nadversarial multivariate time series examples for three state-of-the-art deep\nlearning regression models, specifically Convolutional Neural Network (CNN),\nLong Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). We evaluate our\nstudy using Google stock and household power consumption dataset. The obtained\nresults show that all the evaluated DL regression models are vulnerable to\nadversarial attacks, transferable, and thus can lead to catastrophic\nconsequences in safety-critical and cost-critical domains, such as energy and\nfinance.\n