Synthetic data inherits the differential privacy guarantees of the model used to generate it. Additionally, synthetic data may benefit from privacy amplification when the generative model is kept hidden. While empirical studies suggest this phenomenon, a rigorous theoretical understanding is still lacking. In this paper, we investigate this question through the well-understood framework of linear regression. First, we establish negative results showing that if an adversary controls the seed of the generative model, a single synthetic data point can leak as much information as releasing the model itself. Conversely, we show that when synthetic data is generated from random inputs, releasing a limited number of synthetic data points amplifies privacy beyond the model's inherent guarantees. We believe our findings in linear regression can serve as a foundation for deriving more general bounds in the future.

Pufferfish privacy is a flexible generalization of differential privacy that allows to model arbitrary secrets and adversary’s prior knowledge about the data. Unfortunately, designing general and tractable Pufferfish mechanisms that do not compromise utility is challenging. Furthermore, this framework does not provide the composition guarantees needed for a direct use in iterative machine learning algorithms. To mitigate these issues, we introduce a Rényi divergence-based variant of Pufferfish and show that it allows us to extend the applicability of the Pufferfish framework. We first generalize the Wasserstein mechanism to cover a wide range of noise distributions and introduce several ways to improve its utility. Finally, as an alternative to composition, we prove privacy amplification results for contractive noisy iterations and showcase the first use of Pufferfish in private convex optimization. A common ingredient underlying our results is the use and extension of shift reduction lemmas.

Although machine learning models are found in more and more practical applications, stakeholders can be suspicious about the fact that they are not hard-coded and fully specified. To foster trust, it is crucial to provide models whose predictions are explainable. Decision Trees can be understood by humans if they are simple enough, but they suffer in accuracy when compared to other common machine learning methods. Oblique Decision Trees can provide better accuracy and smaller trees, but their decision rules are more complex. This article presents MUST (Multivariate Understandable Statistical Tree), an Oblique Decision Tree split algorithm based on Linear Discriminant Analysis that aims to preserve explainability by limiting the number of variables that appear in decision rules.

This paper demonstrates the advantages of using structured multimedia documents for session management and media distribution in ubiquitous environments. We show how document manipulations can be used to perform powerful operations such as content to context adaptation and presentation continuity. When consuming media in ubiquitous environments, where the set of devices surrounding a user may change, dynamic media adaptation and session transfer become primary requirements. This paper presents a working system, based on a representative scenario, in which multimedia content is distributed and adapted to a movable user to best suit his/her contextual situation. The implemented scenario includes the following scenes: content selection using a personal mobile phone, content distribution to the most suitable device according to the user's context, and presentation continuity when the user moves to another location. This paper introduces the underlying document manipulations that turn the scenario into a working system.

Unlike mono-agent systems, multi-agent planing addresses the problem of resolving conflicts between individual and group interests. In this paper, we are using a Decentralized Vector Valued Markov Decision Process (2V-DEC-MDP) in order to solve this problem. It uses an utility function which is returning a vector representing both individual and group interest. The individual interest of an agent, computed off-line, is based on its optimal policy. The group interest is computed on-line by the agents using their own local observations. In order to take into account both criteria in a decision process and to find a good trade-off between the group interest and the agent's one, we developed a regret-based algorithm based on the Tchebychev Norm.

This paper introduces a generic architecture that enables the development and execution of mobile multimodal applications proposed within the EU IST-511607 project MobiLife. Mobi Life aims at exploiting the synergetic use of multimodal user interface technology and contextual information processing, with the ultimate goal that the two together can provide a beyond-the-state-of-the-art user experience. And this led to an integrated concept, components of the underlying architecture are described in detail and the interfaces towards the application back-end as well as towards context aware resources are discussed. The paper also positions the current work against existing standardisation efforts and it pinpoints technologies required to support the implementation of a device and modality function within the MobiLife architecture