Self-supervised learning (SSL) has become the de facto training paradigm of large models, where pre-training is followed by supervised fine-tuning using domain-specific data and labels. Despite demonstrating comparable performance with supervised methods, comprehensive efforts to assess SSL's impact on machine learning fairness (i.e., performing equally on different demographic breakdowns) are lacking. Hypothesizing that SSL models would learn more generic, hence less biased representations, this study explores the impact of pre-training and fine-tuning strategies on fairness. We introduce a fairness assessment framework for SSL, comprising five stages: defining dataset requirements, pre-training, fine-tuning with gradual unfreezing, assessing representation similarity conditioned on demographics, and establishing domain-specific evaluation processes. We evaluate our method's generalizability on three real-world human-centric datasets (i.e., MIMIC, MESA, and GLOBEM) by systematically comparing hundreds of SSL and fine-tuned models on various dimensions spanning from the intermediate representations to appropriate evaluation metrics. Our findings demonstrate that SSL can significantly improve model fairness, while maintaining performance on par with supervised methods-exhibiting up to a 30% increase in fairness with minimal loss in performance through self-supervision. We posit that such differences can be attributed to representation dissimilarities found between the best- and the worst-performing demographics across models-up to x13 greater for protected attributes with larger performance discrepancies between segments.

Text watermarking technology aims to tag and identify content produced by large language models (LLMs) to prevent misuse. In this study, we introduce the concept of cross-lingual consistency in text watermarking, which assesses the ability of text watermarks to maintain their effectiveness after being translated into other languages. Preliminary empirical results from two LLMs and three watermarking methods reveal that current text watermarking technologies lack consistency when texts are translated into various languages. Based on this observation, we propose a Cross-lingual Watermark Removal Attack (CWRA) to bypass watermarking by first obtaining a response from an LLM in a pivot language, which is then translated into the target language. CWRA can effectively remove watermarks, decreasing the AUCs to a random-guessing level without performance loss. Furthermore, we analyze two key factors that contribute to the cross-lingual consistency in text watermarking and propose X-SIR as a defense method against CWRA. Code: https://github.com/zwhe99/X-SIR.

We study the application of differential privacy in hyper-parameter tuning, a crucial process in machine learning involving selecting the best hyper-parameter from several candidates. Unlike many private learning algorithms, including the prevalent DP-SGD, the privacy implications of tuning remain insufficiently understood or often totally ignored. Recent works propose a generic private selection solution for the tuning process, yet a fundamental question persists: is this privacy bound tight? This paper provides an in-depth examination of this question. Initially, we provide studies affirming the current privacy analysis for private selection is indeed tight in general. However, when we specifically study the hyper-parameter tuning problem in a white-box setting, such tightness no longer holds. This is first demonstrated by applying privacy audit on the tuning process. Our findings underscore a substantial gap between current theoretical privacy bound and the empirical bound derived even under strong audit setups. This gap motivates our subsequent investigations. Our further study provides improved privacy results for private hyper-parameter tuning due to its distinct properties. Our results demonstrate broader applicability compared to prior analyses, which are limited to specific parameter configurations.

Deep neural networks are highly susceptible to learning biases in visual data. While various methods have been proposed to mitigate such bias, the majority require explicit knowledge of the biases present in the training data in order to mitigate. We argue the relevance of exploring methods which are completely ignorant of the presence of any bias, but are capable of identifying and mitigating them. Furthermore, we propose using Bayesian neural networks with a predictive uncertainty-weighted loss function to dynamically identify potential bias in individual training samples and to weight them during training. We find a positive correlation between samples subject to bias and higher epistemic uncertainties. Finally, we show the method has potential to mitigate visual bias on a bias benchmark dataset and on a real-world face detection problem, and we consider the merits and weaknesses of our approach.

Graph anomaly detection (GAD) is increasingly crucial in various applications, ranging from financial fraud detection to fake news detection. However, current GAD methods largely overlook the fairness problem, which might result in discriminatory decisions skewed toward certain demographic groups defined on sensitive attributes (e.g., gender, religion, ethnicity, etc.). This greatly limits the applicability of these methods in real-world scenarios in light of societal and ethical restrictions. To address this critical gap, we make the first attempt to integrate fairness with utility in GAD decision-making. Specifically, we devise a novel DisEntangle-based FairnEss-aware aNomaly Detection framework on the attributed graph, named DEFEND. DEFEND first introduces disentanglement in GNNs to capture informative yet sensitive-irrelevant node representations, effectively reducing societal bias inherent in graph representation learning. Besides, to alleviate discriminatory bias in evaluating anomalous nodes, DEFEND adopts a reconstruction-based anomaly detection, which concentrates solely on node attributes without incorporating any graph structure. Additionally, given the inherent association between input and sensitive attributes, DEFEND constrains the correlation between the reconstruction error and the predicted sensitive attributes. Our empirical evaluations on real-world datasets reveal that DEFEND performs effectively in GAD and significantly enhances fairness compared to state-of-the-art baselines. To foster reproducibility, our code is available at https://github.com/AhaChang/DEFEND.

Synthetic tabular data generation with differential privacy is a crucial problem to enable data sharing with formal privacy. Despite a rich history of methodological research and development, developing differentially private tabular data generators that can provide realistic synthetic datasets remains challenging. This paper introduces DP-LLMTGen -- a novel framework for differentially private tabular data synthesis that leverages pretrained large language models (LLMs). DP-LLMTGen models sensitive datasets using a two-stage fine-tuning procedure with a novel loss function specifically designed for tabular data. Subsequently, it generates synthetic data through sampling the fine-tuned LLMs. Our empirical evaluation demonstrates that DP-LLMTGen outperforms a variety of existing mechanisms across multiple datasets and privacy settings. Additionally, we conduct an ablation study and several experimental analyses to deepen our understanding of LLMs in addressing this important problem. Finally, we highlight the controllable generation ability of DP-LLMTGen through a fairness-constrained generation setting.

Cross-validation is a common method for estimating the predictive performance of machine learning models. In a data-scarce regime, where one typically wishes to maximize the number of instances used for training the model, an approach called "leave-one-out cross-validation" is often used. In this design, a separate model is built for predicting each data instance after training on all other instances. Since this results in a single test data point available per model trained, predictions are aggregated across the entire dataset to calculate common rank-based performance metrics such as the area under the receiver operating characteristic or precision-recall curves. In this work, we demonstrate that this approach creates a negative correlation between the average label of each training fold and the label of its corresponding test instance, a phenomenon that we term distributional bias. As machine learning models tend to regress to the mean of their training data, this distributional bias tends to negatively impact performance evaluation and hyperparameter optimization. We show that this effect generalizes to leave-P-out cross-validation and persists across a wide range of modeling and evaluation approaches, and that it can lead to a bias against stronger regularization. To address this, we propose a generalizable rebalanced cross-validation approach that corrects for distributional bias. We demonstrate that our approach improves cross-validation performance evaluation in synthetic simulations and in several published leave-one-out analyses.

This paper surveys evaluation techniques to enhance the trustworthiness and understanding of Large Language Models (LLMs). As reliance on LLMs grows, ensuring their reliability, fairness, and transparency is crucial. We explore algorithmic methods and metrics to assess LLM performance, identify weaknesses, and guide development towards more trustworthy applications. Key evaluation metrics include Perplexity Measurement, NLP metrics (BLEU, ROUGE, METEOR, BERTScore, GLEU, Word Error Rate, Character Error Rate), Zero-Shot and Few-Shot Learning Performance, Transfer Learning Evaluation, Adversarial Testing, and Fairness and Bias Evaluation. We introduce innovative approaches like LLMMaps for stratified evaluation, Benchmarking and Leaderboards for competitive assessment, Stratified Analysis for in-depth understanding, Visualization of Blooms Taxonomy for cognitive level accuracy distribution, Hallucination Score for quantifying inaccuracies, Knowledge Stratification Strategy for hierarchical analysis, and Machine Learning Models for Hierarchy Generation. Human Evaluation is highlighted for capturing nuances that automated metrics may miss. These techniques form a framework for evaluating LLMs, aiming to enhance transparency, guide development, and establish user trust. Future papers will describe metric visualization and demonstrate each approach on practical examples.

Effective tracking of surrounding traffic participants allows for an accurate state estimation as a necessary ingredient for prediction of future behavior and therefore adequate planning of the ego vehicle trajectory. One approach for detecting and tracking surrounding traffic participants is the combination of a learning based object detector with a classical tracking algorithm. Learning based object detectors have been shown to work adequately on lidar and camera data, while learning based object detectors using standard radar data input have proven to be inferior. Recently, with the improvements to radar sensor technology in the form of imaging radars, the object detection performance on radar was greatly improved but is still limited compared to lidar sensors due to the sparsity of the radar point cloud. This presents a unique challenge for the task of multi-object tracking. The tracking algorithm must overcome the limited detection quality while generating consistent tracks. To this end, a comparison between different multi-object tracking methods on imaging radar data is required to investigate its potential for downstream tasks. The work at hand compares multiple approaches and analyzes their limitations when applied to imaging radar data. Furthermore, enhancements to the presented approaches in the form of probabilistic association algorithms are considered for this task.

The integration of Large Language Models (LLMs) into healthcare diagnostics offers a promising avenue for clinical decision-making. This study outlines the development of a novel method for zero-shot/few-shot in-context learning (ICL) by integrating medical domain knowledge using a multi-layered structured prompt. We also explore the efficacy of two communication styles between the user and LLMs: the Numerical Conversational (NC) style, which processes data incrementally, and the Natural Language Single-Turn (NL-ST) style, which employs long narrative prompts. Our study systematically evaluates the diagnostic accuracy and risk factors, including gender bias and false negative rates, using a dataset of 920 patient records in various few-shot scenarios. Results indicate that traditional clinical machine learning (ML) models generally outperform LLMs in zero-shot and few-shot settings. However, the performance gap narrows significantly when employing few-shot examples alongside effective explainable AI (XAI) methods as sources of domain knowledge. Moreover, with sufficient time and an increased number of examples, the conversational style (NC) nearly matches the performance of ML models. Most notably, LLMs demonstrate comparable or superior cost-sensitive accuracy relative to ML models. This research confirms that, with appropriate domain knowledge and tailored communication strategies, LLMs can significantly enhance diagnostic processes. The findings highlight the importance of optimizing the number of training examples and communication styles to improve accuracy and reduce biases in LLM applications.