Successfully applying large language models (LLMs) to real-world problems is often best achieved by fine-tuning on domain-specific data (Liu et al., 2022; Mosbach et al., 2023). This approach is seen in a variety of commercial products deployed today, e.g., GitHub Copilot (Chen et al., 2021), Gmail Smart Compose (Chen et al., 2019), GBoard (Xu et al., 2023), etc., that are based on LMs trained or fine-tuned on domain-specific data collected from users. The practice of fine-tuning on user data--particularly on sensitive data like emails, texts, or source code--comes with privacy concerns, as LMs have been shown to leak information from their training data (Carlini et al., 2021), especially as models are scaled larger (Carlini et al., 2023). In this paper, we study the privacy risks posed to users whose data are leveraged to fine-tune LLMs. Most existing privacy attacks on LLMs can be grouped into two categories: membership inference, in which the attacker obtains access to a sample and must determine if it was trained on (Mireshghallah et al., 2022; Mattern et al., 2023; Niu et al., 2023); and extraction attacks, in which the attacker tries to reconstruct the training data by prompting the model with different prefixes (Carlini et al., 2021; Lukas et al., 2023). These threat models make no assumptions about the training data and thus cannot estimate the privacy risk to a user when that user contributes many, likely correlated, training samples. To this end we introduce the novel threat model of user inference, a relevant and realistic privacy attack vector for LLMs fine-tuned on user data, depicted in Figure 1.

Finding suitable health insurance coverage can be challenging for individuals and small enterprises in the USA. The Health Insurance Exchange Public Use Files (Exchange PUFs) dataset provided by CMS offers valuable information on health and dental policies [1]. In this paper, we leverage machine learning algorithms to predict if a health insurance plan covers routine dental services for adults. By analyzing plan type, region, deductibles, out-of-pocket maximums, and copayments, we employ Logistic Regression, Decision Tree, Random Forest, Gradient Boost, Factorization Model and Support Vector Machine algorithms. Our goal is to provide a clinical strategy for individuals and families to select the most suitable insurance plan based on income and expenses.

Class imbalance exists in many classification problems, and since the data is designed for accuracy, imbalance in data classes can lead to classification challenges with a few classes having higher misclassification costs. The Backblaze dataset, a widely used dataset related to hard discs, has a small amount of failure data and a large amount of health data, which exhibits a serious class imbalance. This paper provides a comprehensive overview of research in the field of imbalanced data classification. The discussion is organized into three main aspects: data-level methods, algorithmic-level methods, and hybrid methods. For each type of method, we summarize and analyze the existing problems, algorithmic ideas, strengths, and weaknesses. Additionally, the challenges of unbalanced data classification are discussed, along with strategies to address them. It is convenient for researchers to choose the appropriate method according to their needs.

Deepfakes are synthetic media generated using deep generative algorithms and have posed a severe societal and political threat. Apart from facial manipulation and synthetic voice, recently, a novel kind of deepfakes has emerged with either audio or visual modalities manipulated. In this regard, a new generation of multimodal audio-visual deepfake detectors is being investigated to collectively focus on audio and visual data for multimodal manipulation detection. Existing multimodal (audio-visual) deepfake detectors are often based on the fusion of the audio and visual streams from the video. Existing studies suggest that these multimodal detectors often obtain equivalent performances with unimodal audio and visual deepfake detectors. We conjecture that the heterogeneous nature of the audio and visual signals creates distributional modality gaps and poses a significant challenge to effective fusion and efficient performance. In this paper, we tackle the problem at the representation level to aid the fusion of audio and visual streams for multimodal deepfake detection. Specifically, we propose the joint use of modality (audio and visual) invariant and specific representations. This ensures that the common patterns and patterns specific to each modality representing pristine or fake content are preserved and fused for multimodal deepfake manipulation detection. Our experimental results on FakeAVCeleb and KoDF audio-visual deepfake datasets suggest the enhanced accuracy of our proposed method over SOTA unimodal and multimodal audio-visual deepfake detectors by $17.8$% and $18.4$%, respectively. Thus, obtaining state-of-the-art performance.

Most machine learning (ML) systems assume stationary and matching data distributions during training and deployment. This is often a false assumption. When ML models are deployed on real devices, data distributions often shift over time due to changes in environmental factors, sensor characteristics, and task-of-interest. While it is possible to have a human-in-the-loop to monitor for distribution shifts and engineer new architectures in response to these shifts, such a setup is not cost-effective. Instead, non-stationary automated ML (AutoML) models are needed. This paper presents the Encoder-Adaptor-Reconfigurator (EAR) framework for efficient continual learning under domain shifts. The EAR framework uses a fixed deep neural network (DNN) feature encoder and trains shallow networks on top of the encoder to handle novel data. The EAR framework is capable of 1) detecting when new data is out-of-distribution (OOD) by combining DNNs with hyperdimensional computing (HDC), 2) identifying low-parameter neural adaptors to adapt the model to the OOD data using zero-shot neural architecture search (ZS-NAS), and 3) minimizing catastrophic forgetting on previous tasks by progressively growing the neural architecture as needed and dynamically routing data through the appropriate adaptors and reconfigurators for handling domain-incremental and class-incremental continual learning. We systematically evaluate our approach on several benchmark datasets for domain adaptation and demonstrate strong performance compared to state-of-the-art algorithms for OOD detection and few-/zero-shot NAS.

We study the problem of watermarking large language models (LLMs) generated text -- one of the most promising approaches for addressing the safety challenges of LLM usage. In this paper, we propose a rigorous theoretical framework to quantify the effectiveness and robustness of LLM watermarks. We propose a robust and high-quality watermark method, Unigram-Watermark, by extending an existing approach with a simplified fixed grouping strategy. We prove that our watermark method enjoys guaranteed generation quality, correctness in watermark detection, and is robust against text editing and paraphrasing. Experiments on three varying LLMs and two datasets verify that our Unigram-Watermark achieves superior detection accuracy and comparable generation quality in perplexity, thus promoting the responsible use of LLMs. Code is available at https://github.com/XuandongZhao/Unigram-Watermark.

Seven years ago, researchers proposed a postprocessing method to equalize the error rates of a model across different demographic groups. The work launched hundreds of papers purporting to improve over the postprocessing baseline. We empirically evaluate these claims through thousands of model evaluations on several tabular datasets. We find that the fairness-accuracy Pareto frontier achieved by postprocessing contains all other methods we were feasibly able to evaluate. In doing so, we address two common methodological errors that have confounded previous observations. One relates to the comparison of methods with different unconstrained base models. The other concerns methods achieving different levels of constraint relaxation. At the heart of our study is a simple idea we call unprocessing that roughly corresponds to the inverse of postprocessing. Unprocessing allows for a direct comparison of methods using different underlying models and levels of relaxation.

Fact-checkers are often hampered by the sheer amount of online content that needs to be fact-checked. NLP can help them by retrieving already existing fact-checks relevant to the content being investigated. This paper introduces a new multilingual dataset -- MultiClaim -- for previously fact-checked claim retrieval. We collected 28k posts in 27 languages from social media, 206k fact-checks in 39 languages written by professional fact-checkers, as well as 31k connections between these two groups. This is the most extensive and the most linguistically diverse dataset of this kind to date. We evaluated how different unsupervised methods fare on this dataset and its various dimensions. We show that evaluating such a diverse dataset has its complexities and proper care needs to be taken before interpreting the results. We also evaluated a supervised fine-tuning approach, improving upon the unsupervised method significantly.

As machine learning technology gets applied to actual products and solutions, new challenges have emerged. Models unexpectedly fail to generalize to small changes in the distribution, tend to be confident on novel data they have never seen, or cannot communicate the rationale behind their decisions effectively with the end users. Collectively, we face a trustworthiness issue with the current machine learning technology. This textbook on Trustworthy Machine Learning (TML) covers a theoretical and technical background of four key topics in TML: Out-of-Distribution Generalization, Explainability, Uncertainty Quantification, and Evaluation of Trustworthiness. We discuss important classical and contemporary research papers of the aforementioned fields and uncover and connect their underlying intuitions. The book evolved from the homonymous course at the University of T\"ubingen, first offered in the Winter Semester of 2022/23. It is meant to be a stand-alone product accompanied by code snippets and various pointers to further sources on topics of TML. The dedicated website of the book is https://trustworthyml.io/.

Microsatellite instability-high (MSI-H) is a tumor agnostic biomarker for immune checkpoint inhibitor therapy. However, MSI status is not routinely tested in prostate cancer, in part due to low prevalence and assay cost. As such, prediction of MSI status from hematoxylin and eosin (H&E) stained whole-slide images (WSIs) could identify prostate cancer patients most likely to benefit from confirmatory testing and becoming eligible for immunotherapy. Prostate biopsies and surgical resections from de-identified records of consecutive prostate cancer patients referred to our institution were analyzed. Their MSI status was determined by next generation sequencing. Patients before a cutoff date were split into an algorithm development set (n=4015, MSI-H 1.8%) and a paired validation set (n=173, MSI-H 19.7%) that consisted of two serial sections from each sample, one stained and scanned internally and the other at an external site. Patients after the cutoff date formed the temporal validation set (n=1350, MSI-H 2.3%). Attention-based multiple instance learning models were trained to predict MSI-H from H&E WSIs. The MSI-H predictor achieved area under the receiver operating characteristic curve values of 0.78 (95% CI [0.69-0.86]), 0.72 (95% CI [0.63-0.81]), and 0.72 (95% CI [0.62-0.82]) on the internally prepared, externally prepared, and temporal validation sets, respectively. While MSI-H status is significantly correlated with Gleason score, the model remained predictive within each Gleason score subgroup. In summary, we developed and validated an AI-based MSI-H diagnostic model on a large real-world cohort of routine H&E slides, which effectively generalized to externally stained and scanned samples and a temporally independent validation cohort. This algorithm has the potential to direct prostate cancer patients toward immunotherapy and to identify MSI-H cases secondary to Lynch syndrome.