Yuan Wang

Yuan Wang, Ouxiang Li, Tingting Mu et al.

Recent success of text-to-image (T2I) generation and its increasing practical applications, enabled by diffusion models, require urgent consideration of erasing unwanted concepts, e.g., copyrighted, offensive, and unsafe ones, from the pre-trained models in a precise, timely, and low-cost manner. The twofold demand of concept erasure includes not only a precise removal of the target concept (i.e., erasure efficacy) but also a minimal change on non-target content (i.e., prior preservation), during generation. Existing methods face challenges in maintaining an effective balance between erasure efficacy and prior preservation, and they can be computationally costly. To improve, we propose a precise, fast, and low-cost concept erasure method, called Adaptive Value Decomposer (AdaVD), which is training-free. Our method is grounded in a classical linear algebraic operation of computing the orthogonal complement, implemented in the value space of each cross-attention layer within the UNet of diffusion models. We design a shift factor to adaptively navigate the erasure strength, enhancing effective prior preservation without sacrificing erasure efficacy. Extensive comparative experiments with both training-based and training-free state-of-the-art methods demonstrate that the proposed AdaVD excels in both single and multiple concept erasure, showing 2 to 10 times improvement in prior preservation than the second best, meanwhile achieving the best or near best erasure efficacy. AdaVD supports a series of diffusion models and downstream image generation tasks, with code available on: https://github.com/WYuan1001/AdaVD.

Yuan Wang, Yangxin Fan

We believe that "all men are created equal". With the rise of the police shootings reported by media, more people in the U.S. think that police use excessive force during law enforcement, especially to a specific group of people. We want to apply multidimensional statistical analysis to reveal more facts than the monotone mainstream media. Our paper has three parts. First, we proposed a new method to quantify fatal police shooting news reporting deviation of mainstream media, which includes CNN, FOX, ABC, and NBC. Second, we analyzed the most comprehensive US fatal police shooting dataset from Washington Post. We used FP-growth to reveal the frequent patterns and DBSCAN clustering to find fatal shooting hotspots. We brought multi-attributes (social economics, demographics, political tendency, education, gun ownership rate, police training hours, etc.) to reveal connections under the iceberg. We found that the police shooting rate of a state depends on many variables. The top four most relevant attributes were state joined year, state land area, gun ownership rate, and violent crime rate. Third, we proposed four regression models to predict police shooting rates at the state level. The best model Kstar could predict the fatal police shooting rate with about 88.53% correlation coefficient. We also proposed classification models, including Gradient Boosting Machine, Multi-class Classifier, Logistic Regression, and Naive Bayes Classifier, to predict the race of fatal police shooting victims. Our classification models show no significant evidence to conclude that racial discrimination happened during fatal police shootings recorded by the WP dataset.