Abstract:

Data plays an increasingly crucial role in both the performance and the safety of AI models. Data attribution is an emerging family of techniques aimed at quantifying the impact of individual training data points on a model trained on them, which has found data-centric applications such as training data curation, instance-based explanation, and copyright compensation. In this talk, I will explore practical challenges of deploying data attribution in real-world applications.

In the first part, I will examine the adversarial robustness of data attribution methods, particularly in the context of fairly compensating training data providers. Our study reveals a critical vulnerability, demonstrating how malicious data providers can manipulate these data to unfairly inflate their compensation. In the second part, I will address the limitations in the flexibility of existing influence function approaches and introduce a novel method that extends data attribution to broader machine learning paradigms, including survival analysis and contrastive learning. If time permits, I will also briefly introduce our efforts to tackle challenges related to computational efficiency and group effects in data attribution, and discuss the current advancements and open problems in this field.

Bio:

Jiaqi Ma is an Assistant Professor at the University of Illinois Urbana-Champaign (UIUC). His research interests lie in the broad area of trustworthy AI, with recent focuses including data attribution, machine unlearning, explainable machine learning, and training data curation. Jiaqi’s work has been recognized with the Gary M. Olson Outstanding Student Award from University of Michigan and a Best Paper Award from the DPFM Workshop at ICLR 2024. Prior to joining UIUC, Jiaqi earned his PhD from the University of Michigan and worked as a postdoctoral researcher at Harvard University.

Homepage:  Jiaqi Ma