Adversarial Attacks on Time Series.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Security

Background:

• Time series classification models are increasingly important.
• Generating adversarial samples for these models is an underexplored security concern.
• Adversarial attacks can compromise the integrity of time series data.

Purpose of the Study:

• To propose and evaluate a novel adversarial attack method for time series classification models.
• To assess the vulnerability of common time series classification models to adversarial attacks.
• To demonstrate the effectiveness of the proposed attack compared to existing methods.

Main Methods:

• Utilizing an adversarial transformation network (ATN) on a distilled model as a surrogate.
• Applying the attack methodology to 1-nearest neighbor dynamic time warping (1-NN DTW) and a fully convolutional network (FCN).
• Evaluating the attack's success rate across 42 University of California Riverside (UCR) datasets.

Main Results:

• Both 1-NN DTW and FCN models were found to be susceptible to adversarial attacks on all 42 UCR datasets.
• The proposed ATN-based attack generated a higher fraction of successful black-box attacks compared to the Fast Gradient Sign Method.
• A defense mechanism was successfully developed to mitigate the impact of adversarial samples.

Conclusions:

• Time series classification models are vulnerable to sophisticated adversarial attacks.
• The proposed ATN attack is effective in generating adversarial samples for time series data.
• Incorporating adversarial samples into training datasets is recommended to enhance model resilience.