Poisoning for Debiasing: Fair Recognition via Eliminating Bias Uncovered in Data Poisoning

Yi Zhang, Zhefeng Wang, Rui Hu, Xinyu Duan, Yi ZHENG, Baoxing Huai, Jiarun Han, Jitao Sang

Abstract: Neural networks often tend to rely on bias features that have strong but spurious correlations with the target labels for decision-making, leading to poor performance on data that does not adhere to these correlations. Early debiasing methods typically construct an unbiased optimization objective based on the labels of bias features. Recent work assumes that bias label is unavailable and usually trains two models: a biased model to deliberately learn bias features for exposing data bias, and a target model to eliminate bias captured by the bias model. In this paper, we first reveal that previous biased models fit target labels, which resulted in failing to expose data bias. To tackle this issue, we propose poisoner, which utilizes data poisoning to embed the biases learned by biased models into the poisoned training data, thereby encouraging the models to learn more biases. Specifically, we couple data poisoning and model training to continuously prompt the biased model to learn more bias. By utilizing the biased model, we can identify samples in the data that contradict these biased correlations. Subsequently, we amplify the influence of these samples in the training of the target model to prevent the model from learning such biased correlations. Experiments show the superior debiasing performance of our method.

Primary Subject Area: [Content] Media Interpretation

Secondary Subject Area: [Generation] Social Aspects of Generative AI, [Engagement] Emotional and Social Signals

Relevance To Conference:

Our research aims to enhance the fairness of AI models in the absence of annotations for potential biases. The core challenge lies in understanding potential biases in the data. We propose a novel approach leveraging data poisoning to uncover data biases and subsequently eliminate them. By repurposing the malicious intent behind data poisoning for benign purposes, our methodology uncovers unknown biases and improves fairness.

Our innovative approach of using data poisoning to understand and uncover data biases represents a novel paradigm in content interpretation. This aligns with ACMMM's theme of multimedia interpretation and holds significant implications for promoting a deeper understanding of potential biases in information processing fields. Moreover, our work resonates with this year's ACMMM theme of social impact by offering insights into improving AI fairness, which is a crucial aspect of multimedia processing. Additionally, our exploration of benign applications of data poisoning reflects AI's original intent towards benevolence.

We firmly believe that our work significantly contributes to the advancement of multimedia processing towards more inclusive and unbiased systems, aligning with the goals of the ACMMM community.

Submission Number: 4492