This paper introduces two novel slope-based adversarial attacks—General Slope Attack (GSA) and Least-Squares Slope Attack (LSSA)—designed to manipulate the trend of financial time-series forecasts generated by the N-HiTS model. Unlike traditional attacks that maximize error or shift predictions uniformly, these methods embed a desired slope into the loss function, allowing fine-grained control over temporal dynamics. The GSA alters the slope between endpoints, while the LSSA modifies the overall trend using least-squares regression. Both attacks are evaluated on an N-HiTS model trained on over 300 S&P 500 stocks, demonstrating the ability to double the slope of predictions with minimal perturbation (e.g., 2% of median price). Importantly, these attacks remain stealthy, bypassing a CNN-based discriminator with accuracy as low as 57% and specificity of 28%.

The paper also proposes an Adversarial GAN (A-GAN) that integrates the LSSA loss into a conditional Wasserstein GAN to generate realistic adversarial examples. While the A-GAN successfully increases the slope of N-HiTS predictions, it suffers from mode collapse, making its outputs detectable. Additionally, the authors highlight a critical security gap by presenting a malware injection method that targets the model inference library, enabling gradient-based attacks even when gradients are disabled. This work underscores the importance of securing the entire ML pipeline, not just the model itself, and opens avenues for future research on slope-based attacks in other time-series domains and improved GAN stability.