5 February 2024 in PLOS ONE
Differentially private (DP) synthetic datasets are a solution for sharing data while preserving the privacy of individual data providers. Understanding the effects of utilizing DP synthetic data in end-to-end machine learning pipelines impacts areas such as health care and humanitarian action, where data is scarce and regulated by restrictive privacy laws. In this work, we investigate the extent to which synthetic data can replace real, tabular data in machine learning pipelines and identify the most effective synthetic data generation techniques for training and evaluating machine learning models. We systematically investigate the impacts of differentially private synthetic data on downstream classification tasks from the point of view of utility as well as fairness. Our analysis is comprehensive and includes representatives of the two main types of synthetic data generation algorithms: marginal-based and GAN-based. To the best of our knowledge, our work is the first that: (i) proposes a training and evaluation framework that does not assume that real data is available for testing the utility and fairness of machine learning models trained on synthetic data; (ii) presents the most extensive analysis of synthetic dataset generation algorithms in terms of utility and fairness when used for training machine learning models; and (iii) encompasses several different definitions of fairness. Our findings demonstrate that marginal-based synthetic data generators surpass GAN-based ones regarding model training utility for tabular data. Indeed, we show that models trained using data generated by marginal-based algorithms can exhibit similar utility to models trained using real data. Our analysis also reveals that the marginal-based synthetic data generated using AIM and MWEM PGM algorithms can train models that simultaneously achieve utility and fairness characteristics close to those obtained by models trained with real data.
12 October 2023 in IEEE Transactions on Dependable and Secure Computing
Child Sexual Abuse Media (CSAM) is any visual record of a sexually explicit activity involving minors. Machine learning-based solutions can help law enforcement identify CSAM and block distribution. Yet, collecting CSAM imagery to train machine learning models has ethical and legal constraints. CSAM detection systems based on file metadata offer several opportunities. Metadata is not a record of a crime and, therefore, clear of legal restrictions. This paper proposes a CSAM detection framework consisting of machine learning models trained on file paths extracted from a real-world data set of over 1 million file paths obtained in criminal investigations. Our framework includes guidelines for model evaluation that account for data changes caused by adversarial data modification and variations in data distribution caused by limited access to training data, as well as an assessment of false positive rates against file paths from common crawl data. We achieve accuracies as high as 0.97 while presenting stable behavior under adversarial attacks previously used in natural language tasks. When evaluating the model on publicly available file paths from common crawl data, we observed a false positive rate of 0.002, showing that the model operating in distinct data distributions maintains low false positive rates.
1 November 2024 in Science Advances
Do search engine algorithms systematically expose users to content from unreliable sites? There is widespread concern that they do, but little systematic evidence that search engine algorithms, rather than user-expressed preferences, are driving current exposure to and engagement with unreliable information sources. Using two datasets totaling roughly 14 billion search engine result pages (SERPs) from Bing, the second most popular search engine in the U.S., we show that search exposes users to few unreliable information sources. The vast majority of engagement with unreliable information sources from search occurs when users are explicitly searching for information from those sites, despite those searches being an extremely small share of the overall search volume. Our findings highlight the importance of accounting for user preference when examining engagement with unreliable sources from web search.