Barry Chen

Rafael Rivera Soto, Kailin Koch, Aleem Khan et al.

The advent of instruction-tuned language models that convincingly mimic human writing poses a significant risk of abuse. However, such abuse may be counteracted with the ability to detect whether a piece of text was composed by a language model rather than a human author. Some previous approaches to this problem have relied on supervised methods by training on corpora of confirmed human- and machine- written documents. Unfortunately, model under-specification poses an unavoidable challenge for neural network-based detectors, making them brittle in the face of data shifts, such as the release of newer language models producing still more fluent text than the models used to train the detectors. Other approaches require access to the models that may have generated a document in question, which is often impractical. In light of these challenges, we pursue a fundamentally different approach not relying on samples from language models of concern at training time. Instead, we propose to leverage representations of writing style estimated from human-authored text. Indeed, we find that features effective at distinguishing among human authors are also effective at distinguishing human from machine authors, including state-of-the-art large language models like Llama-2, ChatGPT, and GPT-4. Furthermore, given a handful of examples composed by each of several specific language models of interest, our approach affords the ability to predict which model generated a given document. The code and data to reproduce our experiments are available at https://github.com/LLNL/LUAR/tree/main/fewshot_iclr2024.

Karl Ni, Kyle Zaragoza, Charles Foster et al.

Traditional image tagging and retrieval algorithms have limited value as a result of being trained with heavily curated datasets. These limitations are most evident when arbitrary search words are used that do not intersect with training set labels. Weak labels from user generated content (UGC) found in the wild (e.g., Google Photos, FlickR, etc.) have an almost unlimited number of unique words in the metadata tags. Prior work on word embeddings successfully leveraged unstructured text with large vocabularies, and our proposed method seeks to apply similar cost functions to open source imagery. Specifically, we train a deep learning image tagging and retrieval system on large scale, user generated content (UGC) using sampling methods and joint optimization of word embeddings. By using the Yahoo! FlickR Creative Commons (YFCC100M) dataset, such an approach builds robustness to common unstructured data issues that include but are not limited to irrelevant tags, misspellings, multiple languages, polysemy, and tag imbalance. As a result, the final proposed algorithm will not only yield comparable results to state of the art in conventional image tagging, but will enable new capability to train algorithms on large, scale unstructured text in the YFCC100M dataset and outperform cited work in zero-shot capability.

Rafael Rivera Soto, Barry Chen, Nicholas Andrews

Despite considerable progress in the development of machine-text detectors, it has been suggested that the problem is inherently hard, and therefore, that stakeholders should proceed under the assumption that machine-generated text cannot be reliably detected as such. We examine a recent such claim by Nicks et al. (2024) regarding the ease with which language models can be optimized to degrade the performance of machine-text detectors, including detectors not specifically optimized against. We identify a feature space -- the stylistic feature space -- that is robust to such optimization, and show that it may be used to reliably detect samples from language models optimized to prevent detection. Furthermore, we show that even when models are explicitly optimized against stylistic detectors, detection performance remains surprisingly unaffected. We then seek to understand if stylistic detectors are inherently more robust. To study this question, we explore a new paraphrasing approach that simultaneously aims to close the gap between human writing and machine writing in stylistic feature space while avoiding detection using traditional features. We show that when only a single sample is available for detection, this attack is universally effective across all detectors considered, including those that use writing style. However, as the number of samples available for detection grows, the human and machine distributions become distinguishable. Overall, our findings underscore previous recommendations to avoid reliance on machine-text detection.

Rafael Rivera Soto, Barry Chen, Nicholas Andrews

High-quality paraphrases are easy to produce using instruction-tuned language models or specialized paraphrasing models. Although this capability has a variety of benign applications, paraphrasing attacks$\unicode{x2013}$paraphrases applied to machine-generated texts$\unicode{x2013}$are known to significantly degrade the performance of machine-text detectors. This motivates us to consider the novel problem of paraphrase inversion, where, given paraphrased text, the objective is to recover an approximation of the original text. The closer the approximation is to the original text, the better machine-text detectors will perform. We propose an approach which frames the problem as translation from paraphrased text back to the original text, which requires examples of texts and corresponding paraphrases to train the inversion model. Fortunately, such training data can easily be generated, given a corpus of original texts and one or more paraphrasing models. We find that language models such as GPT-4 and Llama-3 exhibit biases when paraphrasing which an inversion model can learn with a modest amount of data. Perhaps surprisingly, we also find that such models generalize well, including to paraphrase models unseen at training time. Finally, we show that when combined with a paraphrased-text detector, our inversion models provide an effective defense against paraphrasing attacks, and overall our approach yields an average improvement of +22% AUROC across seven machine-text detectors and three different domains.

Karl Ni, Roger Pearce, Kofi Boakye et al.

We present a work-in-progress snapshot of learning with a 15 billion parameter deep learning network on HPC architectures applied to the largest publicly available natural image and video dataset released to-date. Recent advancements in unsupervised deep neural networks suggest that scaling up such networks in both model and training dataset size can yield significant improvements in the learning of concepts at the highest layers. We train our three-layer deep neural network on the Yahoo! Flickr Creative Commons 100M dataset. The dataset comprises approximately 99.2 million images and 800,000 user-created videos from Yahoo's Flickr image and video sharing platform. Training of our network takes eight days on 98 GPU nodes at the High Performance Computing Center at Lawrence Livermore National Laboratory. Encouraging preliminary results and future research directions are presented and discussed.