Abstract
Predicting the trajectories of surrounding objects is a critical task in self-driving and many other autonomous systems. Recent works demonstrate that adversarial attacks on trajectory prediction, where small crafted perturbations are introduced to history trajectories, may significantly mislead the prediction of future trajectories and ultimately induce unsafe planning. However, few works have addressed enhancing the robustness of this important safety-critical task. In this paper, we present the first adversarial training method for trajectory prediction. Compared with typical adversarial training on image tasks, our work is challenged by more random inputs with rich context, and a lack of class labels. To address these challenges, we propose a method based on a semi-supervised adversarial autoencoder that models disentangled semantic features with domain knowledge and provides additional latent labels for the adversarial training. Extensive experiments with different types of attacks demonstrate that our semi-supervised semantics-guided adversarial training method can effectively mitigate the impact of adversarial attacks and generally improve the system's adversarial robustness to a variety of attacks, including unseen ones. We believe that such semantics-guided architecture and advancement in robust generalization is an important step for developing robust prediction models and enabling safe decision making.
Introduction
In the architecture of autonomous vehicles (AVs), the prediction module serves as a critical "crystal ball," analyzing the historical movement of surrounding agents to forecast their future positions. While deep learning has significantly advanced the accuracy of these predictions, recent research reveals a startling vulnerability: adversarial attacks. By introducing small, "crafted" perturbations to a vehicle's historical trajectory—often unnoticeable to human observers—an attacker can trick an AV's AI into predicting non-existent collisions or missing real ones. Such failures can trigger dangerous emergency maneuvers or lead to catastrophic planning decisions.
While adversarial training (AT) is a standard defense in image recognition, it is difficult to apply to trajectory prediction. Trajectories are continuous time-series data lacking discrete class labels (like "cat" or "dog"), and standard AT often suffers from poor generalization, where a model becomes robust to one specific attack but remains vulnerable to others. This paper introduces Semi-supervised Semantics-guided Adversarial Training (SSAT), the first framework designed to bake "semantic common sense" into trajectory models to enhance their adversarial robustness.
The SSAT Framework
The core innovation of SSAT lies in its structured internal representation. Rather than treating trajectory data as a black-box vector, the method uses a Semi-supervised Adversarial Autoencoder (AAE) to disentangle the latent space into three distinct semantic components:
- Longitudinal Features: These model speed and "time headway" (the gap between vehicles).
- Lateral Features: These capture intent, such as moving forward, turning left, or turning right.
- Residual Features: These capture the remaining stochastic noise and fine-grained details.
Semi-supervised Learning and Adversarial Pipeline
By leveraging the ground truth of what a vehicle actually did during training (e.g., "the car turned left"), the model is forced to map its internal representations to these human-understandable semantic labels. This semi-supervised approach ensures the model understands the intent behind a movement.
During the training process, the system generates adversarial examples using Projected Gradient Descent (PGD). When an attack successfully misleads the model, SSAT employs a weighted loss function that prioritizes maintaining correct high-level semantics before addressing fine-grained coordinate accuracy. Additionally, the framework integrates MixUp, a technique that blends adversarial and benign scenarios to balance the trade-off between robustness and standard accuracy.
Experimental Evaluation
The researchers evaluated SSAT using three industry-standard benchmarks:
- Argoverse 1 & 2: Featuring over 250,000 real-world driving scenarios from complex urban environments like Miami and Pittsburgh.
- ApolloScape: A large-scale dataset providing diverse trajectory patterns.
SSAT was tested against various attack vectors, including those targeting Average Displacement Error (ADE) and specific lateral deviations.
Results and Discussion
SSAT demonstrated significant improvements in adversarial robustness across all datasets:
- Substantial Error Reduction: Compared to unprotected models, SSAT reduced prediction errors under attack by 32% to 73%.
- Superior Generalization: A major breakthrough of this work is its ability to defend against unseen attacks. While standard adversarial training often fails when the attack type changes (e.g., training on random noise but testing on lateral shifts), SSAT maintained high performance because it learned the underlying physics and intent of driving.
- The Denoising Effect: The encoder-decoder architecture effectively acts as a semantic filter, "denoising" high-frequency adversarial jitters before they can influence the final prediction.
Limitations
As is common in adversarial defense, there is a "no free lunch" trade-off. The AAE architecture introduces a slight reduction in accuracy on benign (non-adversarial) data. This is attributed to the "information bottleneck" created by forcing the model to compress complex trajectories into specific semantic latent variables.
Conclusion
SSAT represents a significant step toward robust trajectory prediction in safety-critical autonomous systems. By guiding adversarial training with semantic domain knowledge, the model moves beyond simple pattern matching to a deeper understanding of vehicle behavior. Future work will explore more complex semantic labels—such as "yielding" or "aggressive driving"—and investigate the application of this semantics-guided philosophy to the planning and decision-making modules of the autonomous stack.
Summary updated