Vulnerabilities in Trajectory Prediction Models: A Threat to Autonomous Driving

The study explores the vulnerability of deep learning models, particularly those used in computer vision tasks, to "dangerous correlations" during training. These correlations can be introduced unintentionally through various means, such as noisy data or biased training examples. The researchers investigate the impact of these correlations on model performance and find that they can lead to undesirable outcomes, including propaganda and backdoors.
To better understand the issue, the authors analyze several popular deep learning models used in computer vision tasks and show that even small amounts of noisy data during training can result in significant changes to the model’s behavior. They also demonstrate that these changes are not limited to specific tasks but can have broader implications for other applications as well.
The study highlights the importance of carefully evaluating the training process to avoid introducing dangerous correlations into deep learning models. The authors propose a method called "strong data augmentation" to mitigate this issue, which involves increasing the diversity of the training data to reduce the likelihood of encountering noisy examples.
In summary, the article raises concerns about the potential risks associated with deep learning models and highlights the need for caution when training these models. The authors propose a simple yet effective method to mitigate these risks, which can help maintain the reliability and safety of these models in various applications.