Colliding with Adversaries at ECML-PKDD 2025 Model Robustness Competition 1st Prize Solution

2025-10-18

AI Summary

Addressing the robustness of binary classification under Random Distribution Shuffling Attacks (RDSA) in high-energy physics, this paper proposes a novel artificial neural network architecture: a shared-weight feature embedding module unifies heterogeneous input types, while a densely fused deep fully connected tail enhances perturbation-invariant representation learning. Additionally, we introduce the first million-scale RDSA adversarial sample generation method, enabling end-to-end robust training. Evaluated on the ECML-PKDD 2025 competition task, our model achieves 80.0% accuracy on mixed clean and adversarial data—outperforming the second-best method by +2.0 percentage points and securing first place. This work establishes a scalable architectural paradigm and a standardized data-generation benchmark for out-of-distribution robust learning in physics discovery scenarios.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Multiagent Systems: Adversarial Agents

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

This report presents the winning solution for Task 2 of Colliding with Adversaries: A Challenge on Robust Learning in High Energy Physics Discovery at ECML-PKDD 2025. The goal of the challenge was to design and train a robust ANN-based model capable of achieving high accuracy in a binary classification task on both clean and adversarial data generated with the Random Distribution Shuffle Attack (RDSA). Our solution consists of two components: a data generation phase and a robust model training phase. In the first phase, we produced 15 million artificial training samples using a custom methodology derived from Random Distribution Shuffle Attack (RDSA). In the second phase, we introduced a robust architecture comprising (i)a Feature Embedding Block with shared weights among features of the same type and (ii)a Dense Fusion Tail responsible for the final prediction. Training this architecture on our adversarial dataset achieved a mixed accuracy score of 80%, exceeding the second-place solution by two percentage points.

Problem

• Design robust ANN model for adversarial data classification
• Achieve high accuracy on clean and adversarial RDSA datasets
• Improve model robustness using custom adversarial training methodology

Innovation

• Generated 15 million adversarial samples using custom RDSA methodology
• Used shared-weight Feature Embedding Block for similar features
• Implemented Dense Fusion Tail architecture for final predictions