Fooling Neural Networks for Motion Forecasting via Adversarial Attacks

TL;DR

The results suggest that models are susceptible to attacks even on low levels of perturbation, similar to earlier CNN models, motion forecasting tasks are susceptible to small perturbations and simple 3D transformations.

Abstract

Human motion prediction is still an open problem, which is extremely important for autonomous driving and safety applications. Although there are great advances in this area, the widely studied topic of adversarial attacks has not been applied to multi-regression models such as GCNs and MLP-based architectures in human motion prediction. This work intends to reduce this gap using extensive quantitative and qualitative experiments in state-of-the-art architectures similar to the initial stages of adversarial attacks in image classification. The results suggest that models are susceptible to attacks even on low levels of perturbation. We also show experiments with 3D transformations that affect the model performance, in particular, we show that most models are sensitive to simple rotations and translations which do not alter joint distances. We conclude that similar to earlier CNN models, motion forecasting tasks are susceptible to small perturbations and simple 3D transformations.

Figures (8)

• Figure 1: General diagram block for our experiments.
• Figure 2: Visual interpretation of an algorithm for regression tasks.
• Figure 3: Result of FGSM with increasing epsilon on average MPJPE.
• Figure 4: Result of MIFGSM on average MPJPE with $\mu$ ranging from 0.0 to 2.0 with granularity 0.1 using 10 iterations.
• Figure 5: Transformation effects on the test set using the average MPJPE over the 25 output frames