The increased screen time and isolation caused by the COVID-19 pandemic have led to a significant surge in cases of online grooming, which is the use of strategies by predators to lure children into sexual exploitation. Previous efforts to detect grooming in industry and academia have involved accessing and monitoring private conversations through centrally-trained models or sending private conversations to a global server. In this work, we implement a privacy-preserving pipeline for the early detection of sexual predators. We leverage federated learning and differential privacy in order to create safer online spaces for children while respecting their privacy. We investigate various privacy-preserving implementations and discuss their benefits and shortcomings. Our extensive evaluation using real-world data proves that privacy and utility can coexist with only a slight reduction in utility.

In this paper, we introduce WaKA (Wasserstein K-nearest neighbors Attribution), a novel attribution method that leverages principles from the LiRA (Likelihood Ratio Attack) framework and k-nearest neighbors classifiers (k-NN). WaKA efficiently measures the contribution of individual data points to the model’s loss distribution, analyzing every possible k-NN that can be constructed using the training set, without requiring to sample subsets of the training set. WaKA is versatile and can be used a posteriori as a membership inference attack (MIA) to assess privacy risks or a priori for privacy influence measurement and data valuation. Thus, WaKA can be seen as bridging the gap between data attribution and membership inference attack (MIA) by providing a unified framework to distinguish between a data point’s value and its privacy risk. For instance, we have shown that self-attribution values are more strongly correlated with the attack success rate than the contribution of a point to the model generalization. WaKA’s different usages were also evaluated across diverse real-world datasets, demonstrating performance very close to LiRA when used as an MIA on k-NN classifiers, but with greater computational efficiency. Additionally, WaKA shows greater robustness than Shapley Values for data minimization tasks (removal or addition) on imbalanced datasets.

Opinion mining aims to detect and extract relevant information from large quantity of customer reviews. Automatic opinion summarization then seeks to create a consensual point of view often oriented toward the main sentiment of clients to render their experience. Although factual information is valuable for companies to understand what works or not in their products, summarization approaches that convey objectivity, and constructive feedback from customer reviews have yet to be explored. We propose an adversarial multi-task learning model for document summarization to address this new issue. Our algorithm combines an autoencoder for document summarization with a gradient reversal layer to learn independent representations of subjective and sentiment-based material. We assess and compare our method on the Amazon product review dataset where we introduce an original evaluation dataset for objective summarization. We further completed the analysis with neutrality and objectivity metrics. This study demonstrates that the generated summaries carry out relevant and objective content but also emphasize the importance of various processes and layers in multi-task learners to control the information effectively.

A long-running event represents a continuous stream of information on a given topic, such as natural disasters, stock market updates, or even ongoing customer relationship. These news stories include hundreds of individual, time-dependent texts. Simultaneously, new technologies have profoundly transformed the way we consume information. The need to obtain quick, relevant, and digest updates continuously has become a crucial issue and creates new challenges for the task of automatic document summarization. To that end, we introduce an innovative unsupervised method based on two competing sequence-to-sequence models to produce short updated summaries. The proposed architecture relies on several parameters to balance the outputs from the two autoencoders. This relation enables the overall model to correlate generated summaries with relevant information coming from both current and previous news iterations. Depending on the model configuration, we are then able to control the novelty or the consistency of terms included in generated summaries. We evaluate our method on a modified version of the TREC 2013, 2014, and 2015 datasets to track continuous events from a single source. We not only achieve state-of-the-art performance similar to other more complex unsupervised sentence compression approaches, but also influence the information included in the model in the summaries.