As the quality of image generators continues to improve, deepfakes become a topic of considerable societal debate. Image watermarking allows responsible model owners to detect and label their AI-generated content, which can mitigate the harm. Yet, current state-of-the-art methods in image watermarking remain vulnerable to forgery and removal attacks. This vulnerability occurs in part because watermarks distort the distribution of generated images, unintentionally revealing information about the watermarking techniques. In this work, we first demonstrate a distortion-free watermarking method for images, based on a diffusion model's initial noise. However, detecting the watermark requires comparing the initial noise reconstructed for an image to all previously used initial noises. To mitigate these issues, we propose a two-stage watermarking framework for efficient detection. During generation, we augment the initial noise with generated Fourier patterns to embed information about the group of initial noises we used. For detection, we (i) retrieve the relevant group of noises, and (ii) search within the given group for an initial noise that might match our image. This watermarking approach achieves state-of-the-art robustness to forgery and removal against a large battery of attacks.

The presented report evaluates Contextualizing Hate Speech Classifiers with Post-hoc Explanation Kennedy et al. (2020) paper within the scope of ML Reproducibility Challenge 2020. Our work focuses on both aspects constituting the paper: the method itself and the validity of the stated results. In the following sections, we have described the paper, related works, algorithmic frameworks, our experiments and evaluations. Scope of Reproducibility For the GHC (a dataset), the most important difference between BERT WR and BERT SOC is the increase in recall. While, for Stormfront (a dataset), there are similar improvements for in-domain data and the NYT dataset. But, for verifying the claims we also have tried to run the same experiment on a new data-set.

Prediction bias in machine learning models refers to unintended model behaviors that discriminate against inputs mentioning or produced by certain groups; for example, hate speech classifiers predict more false positives for neutral text mentioning specific social groups. Mitigating bias for each task or domain is inefficient, as it requires repetitive model training, data annotation (e.g., demographic information), and evaluation. In pursuit of a more accessible solution, we propose the Upstream Bias Mitigation for Downstream Fine-Tuning (UBM) framework, which mitigate one or multiple bias factors in downstream classifiers by transfer learning from an upstream model. In the upstream bias mitigation stage, explanation regularization and adversarial training are applied to mitigate multiple bias factors. In the downstream fine-tuning stage, the classifier layer of the model is re-initialized, and the entire model is fine-tuned to downstream tasks in potentially novel domains without any further bias mitigation. We expect downstream classifiers to be less biased by transfer learning from de-biased upstream models. We conduct extensive experiments varying the similarity between the source and target data, as well as varying the number of dimensions of bias (e.g., discrimination against specific social groups or dialects). Our results indicate the proposed UBM framework can effectively reduce bias in downstream classifiers.

A team of USC researchers has created a hate speech classifier that is more context-sensitive, and less likely to mistake a post containing a group identifier as hate speech. Understanding what makes something harmful or offensive can be hard enough for humans, never mind artificial intelligence systems. So, perhaps it's no surprise that social media hate speech detection algorithms, designed to stop the spread of hateful speech, can actually amplify racial bias by blocking inoffensive tweets by black people or other minority group members. In fact, one previous study showed that AI models were 1.5 times more likely to flag tweets written by African Americans as "offensive"--in other words, a false positive--compared to other tweets. Because the current automatic detection models miss out on something vital: context.