Online social media is rife with offensive and hateful comments, prompting the need for their automatic detection given the sheer amount of posts created every second. Creating high-quality human-labelled datasets for this task is difficult and costly, especially because non-offensive posts are significantly more frequent than offensive ones. However, unlabelled data is abundant, easier, and cheaper to obtain. In this scenario, self-training methods, using weakly-labelled examples to increase the amount of training data, can be employed. Recent "noisy" self-training approaches incorporate data augmentation techniques to ensure prediction consistency and increase robustness against noisy data and adversarial attacks. In this paper, we experiment with default and noisy self-training using three different textual data augmentation techniques across five different pre-trained BERT architectures varying in size. We evaluate our experiments on two offensive/hate-speech datasets and demonstrate that (i) self-training consistently improves performance regardless of model size, resulting in up to +1.5% F1-macro on both datasets, and (ii) noisy self-training with textual data augmentations, despite being successfully applied in similar settings, decreases performance on offensive and hate-speech domains when compared to the default method, even with state-of-the-art augmentations such as backtranslation.

Large language models (LLMs) are currently at the forefront of intertwining artificial intelligence (AI) systems with human communication and everyday life. Thus, aligning them with human values is of great importance. However, given the steady increase in reasoning abilities, future LLMs are under suspicion of becoming able to deceive human operators and utilizing this ability to bypass monitoring efforts. As a prerequisite to this, LLMs need to possess a conceptual understanding of deception strategies. This study reveals that such strategies emerged in state-of-the-art LLMs, such as GPT-4, but were non-existent in earlier LLMs. We conduct a series of experiments showing that state-of-the-art LLMs are able to understand and induce false beliefs in other agents, that their performance in complex deception scenarios can be amplified utilizing chain-of-thought reasoning, and that eliciting Machiavellianism in LLMs can alter their propensity to deceive. In sum, revealing hitherto unknown machine behavior in LLMs, our study contributes to the nascent field of machine psychology.

In this work, we present a new theoretical framework to measure re-identification risk in such user representations. Our framework, based on hypothesis testing, formally bounds the probability that an attacker may be able to obtain the identity of a user from their representation. As an application, we show how our framework is general enough to model important real-world applications such as the Chrome's Topics API for interest-based advertising. We complement our theoretical bounds by showing provably good attack algorithms for re-identification that we use to estimate the re-identification risk in the Topics API. We believe this work provides a rigorous and interpretable notion of re-identification risk and a framework to measure it that can be used to inform real-world applications.

We introduce DiffOpt.jl, a Julia library to differentiate through the solution of optimization problems with respect to arbitrary parameters present in the objective and/or constraints. The library builds upon MathOptInterface, thus leveraging the rich ecosystem of solvers and composing well with modeling languages like JuMP. DiffOpt offers both forward and reverse differentiation modes, enabling multiple use cases from hyperparameter optimization to backpropagation and sensitivity analysis, bridging constrained optimization with end-to-end differentiable programming. DiffOpt is built on two known rules for differentiating quadratic programming and conic programming standard forms. However, thanks ability to differentiate through model transformation, the user is not limited to these forms and can differentiate with respect to the parameters of any model that can be reformulated into these standard forms. This notably includes programs mixing affine conic constraints and convex quadratic constraints or objective function.

Machine learning practitioners often fine-tune generative pre-trained models like GPT-3 to improve model performance at specific tasks. Previous works, however, suggest that fine-tuned machine learning models memorize and emit sensitive information from the original fine-tuning dataset. Companies such as OpenAI offer fine-tuning services for their models, but no prior work has conducted a memorization attack on any closed-source models. In this work, we simulate a privacy attack on GPT-3 using OpenAI's fine-tuning API. Our objective is to determine if personally identifiable information (PII) can be extracted from this model. We (1) explore the use of naive prompting methods on a GPT-3 fine-tuned classification model, and (2) we design a practical word generation task called Autocomplete to investigate the extent of PII memorization in fine-tuned GPT-3 within a real-world context. Our findings reveal that fine-tuning GPT3 for both tasks led to the model memorizing and disclosing critical personally identifiable information (PII) obtained from the underlying fine-tuning dataset. To encourage further research, we have made our codes and datasets publicly available on GitHub at: https://github.com/albertsun1/gpt3-pii-attacks

In recent years, self-supervised learning (SSL) models have produced promising results in a variety of speech-processing tasks, especially in contexts of data scarcity. In this paper, we study the use of SSL models for the task of mispronunciation detection for second language learners. We compare two downstream approaches: 1) training the model for phone recognition (PR) using native English data, and 2) training a model directly for the target task using non-native English data. We compare the performance of these two approaches for various SSL representations as well as a representation extracted from a traditional DNN-based speech recognition model. We evaluate the models on L2Arctic and EpaDB, two datasets of non-native speech annotated with pronunciation labels at the phone level. Overall, we find that using a downstream model trained for the target task gives the best performance and that most upstream models perform similarly for the task.

In the digital age, data is a valuable commodity, and data marketplaces offer lucrative opportunities for data owners to monetize their private data. However, data privacy is a significant concern, and differential privacy has become a popular solution to address this issue. Private data trading systems (PDQS) facilitate the trade of private data by determining which data owners to purchase data from, the amount of privacy purchased, and providing specific aggregation statistics while protecting the privacy of data owners. However, existing PDQS with separated procurement and query processes are prone to over-perturbation of private data and lack trustworthiness. To address this issue, this paper proposes a framework for PDQS with an integrated procurement and query process to avoid excessive perturbation of private data. We also present two instances of this framework, one based on a greedy approach and another based on a neural network. Our experimental results show that both of our mechanisms outperformed the separately conducted procurement and query mechanism under the same budget regarding accuracy.

Data augmentation is classically used to improve the overall performance of deep learning models. It is, however, challenging in the case of medical applications, and in particular for multiparametric datasets. For example, traditional geometric transformations used in several applications to generate synthetic images can modify in a non-realistic manner the patients' anatomy. Therefore, dedicated image generation techniques are necessary in the medical field to, for example, mimic a given pathology realistically. This paper introduces a new data augmentation architecture that generates synthetic multiparametric (T1 arterial, T1 portal, and T2) magnetic resonance images (MRI) of massive macrotrabecular subtype hepatocellular carcinoma with their corresponding tumor masks through a generative deep learning approach. The proposed architecture creates liver tumor masks and abdominal edges used as input in a Pix2Pix network for synthetic data creation. The method's efficiency is demonstrated by training it on a limited multiparametric dataset of MRI triplets from $89$ patients with liver lesions to generate $1,000$ synthetic triplets and their corresponding liver tumor masks. The resulting Frechet Inception Distance score was $86.55$. The proposed approach was among the winners of the 2021 data augmentation challenge organized by the French Society of Radiology.

This paper presents a classification model that predicts delays in Indian lower courts based on case information available at filing. The model is built on a dataset of 4.2 million court cases filed in 2010 and their outcomes over a 10-year period. The data set is drawn from 7000+ lower courts in India. The authors employed AutoML to develop a multi-class classification model over all periods of pendency and then used binary decision forest classifiers to improve predictive accuracy for the classification of delays. The best model achieved an accuracy of 81.4%, and the precision, recall, and F1 were found to be 0.81. The study demonstrates the feasibility of AI models for predicting delays in Indian courts, based on relevant data points such as jurisdiction, court, judge, subject, and the parties involved. The paper also discusses the results in light of relevant literature and suggests areas for improvement and future research. The authors have made the dataset and Python code files used for the analysis available for further research in the crucial and contemporary field of Indian judicial reform.

One of the key AI tools for textual corpora exploration is natural language question-answering (QA). Unlike keyword-based search engines, QA algorithms receive and process natural language questions and produce precise answers to these questions, rather than long lists of documents that need to be manually scanned by the users. State-of-the-art QA algorithms based on DNNs were successfully employed in various domains. However, QA in the genealogical domain is still underexplored, while researchers in this field (and other fields in humanities and social sciences) can highly benefit from the ability to ask questions in natural language, receive concrete answers and gain insights hidden within large corpora. While some research has been recently conducted for factual QA in the genealogical domain, to the best of our knowledge, there is no previous research on the more challenging task of numerical aggregation QA (i.e., answering questions combining aggregation functions, e.g., count, average, max). Numerical aggregation QA is critical for distant reading and analysis for researchers (and the general public) interested in investigating cultural heritage domains. Therefore, in this study, we present a new end-to-end methodology for numerical aggregation QA for genealogical trees that includes: 1) an automatic method for training dataset generation; 2) a transformer-based table selection method, and 3) an optimized transformer-based numerical aggregation QA model. The findings indicate that the proposed architecture, GLOBE, outperforms the state-of-the-art models and pipelines by achieving 87% accuracy for this task compared to only 21% by current state-of-the-art models. This study may have practical implications for genealogical information centers and museums, making genealogical data research easy and scalable for experts as well as the general public.