Uncertainty Quantification for Collaborative Object Detection Under Adversarial Attacks

TL;DR

The paper tackles uncertainty quantification for Collaborative Object Detection (COD) under adversarial attacks in multi-agent perception. It proposes TUQCP, which combines white-box adversarial training on shared information with a learning-based UQ module and conformal prediction to provide calibrated uncertainty estimates, optimizing a joint loss $\mathcal{L}=w_1\mathcal{L}_{reg}+w_2\mathcal{L}_{cls}+w_3\mathcal{L}_{UQ}$. By simulating attacks via Projected Gradient Descent (PGD) on shared data, and calibrating output uncertainty with conformal prediction, TUQCP achieves strong robustness gains, including an average $80.41\%$ improvement in object detection accuracy over baselines under the same attacks and reduced uncertainty metrics like $\text{KLD}$ and $\text{NLL}$. The approach is demonstrated on the V2X-Sim dataset and is compatible with early, intermediate, and single-agent COD models, offering a practical pathway toward reliable, credible COD in autonomous driving and related domains.

Abstract

Collaborative Object Detection (COD) and collaborative perception can integrate data or features from various entities, and improve object detection accuracy compared with individual perception. However, adversarial attacks pose a potential threat to the deep learning COD models, and introduce high output uncertainty. With unknown attack models, it becomes even more challenging to improve COD resiliency and quantify the output uncertainty for highly dynamic perception scenes such as autonomous vehicles. In this study, we propose the Trusted Uncertainty Quantification in Collaborative Perception framework (TUQCP). TUQCP leverages both adversarial training and uncertainty quantification techniques to enhance the adversarial robustness of existing COD models. More specifically, TUQCP first adds perturbations to the shared information of randomly selected agents during object detection collaboration by adversarial training. TUQCP then alleviates the impacts of adversarial attacks by providing output uncertainty estimation through learning-based module and uncertainty calibration through conformal prediction. Our framework works for early and intermediate collaboration COD models and single-agent object detection models. We evaluate TUQCP on V2X-Sim, a comprehensive collaborative perception dataset for autonomous driving, and demonstrate a 80.41% improvement in object detection accuracy compared to the baselines under the same adversarial attacks. TUQCP demonstrates the importance of uncertainty quantification to COD under adversarial attacks.

Figures (2)

• Figure 1: TUQCP randomly selects $M$ agents as attackers and generates minimum perturbation ${\delta}_m$ to the shared information $F_m$ of each attacker to the ego-agent, such thatthe object detection error is maximized. Besides predicting the classification probability $\hat{p}$ and the location $\hat{y}$ of targeted objects by the COD model, TUQCP quantifies the preliminary object detection uncertainty $\hat{p}$ of each object through the proposed learning-based UQ (DM) component and calibrates the uncertainty through the proposed conformal prediction (CP) component by the conformal quantile $\hat{q}$.
• Figure 2: Visualization of the object detection results of base model (DiscoNet (PGD Test)) and DiscoNet enhanced with our TUQCP (DiscoNet+ TUQCP), both under the same adversarial attack in the testing stage. Red boxes are predictions, and green boxes are ground truth.