Lorenzo Rossi

Paul Égré, Lorenzo Rossi, Jan Sprenger

This paper develops a trivalent semantics for the truth conditions and the probability of the natural language indicative conditional. Our framework rests on trivalent truth conditions first proposed by W. Cooper and yields two logics of conditional reasoning: (i) a logic C of inference from certain premises; and (ii) a logic U of inference from uncertain premises. But whereas C is monotonic for the conditional, U is not, and whereas C obeys Modus Ponens, U does not without restrictions. We show systematic correspondences between trivalent and probabilistic representations of inferences in either framework, and we use the distinction between the two systems to cast light, in particular, on McGee's puzzle about Modus Ponens. The result is a unified account of the semantics and epistemology of indicative conditionals that can be fruitfully applied to analyzing the validity of conditional inferences.

Lorenzo Rossi

Deep neural networks and other machine learning systems, despite being extremely powerful and able to make predictions with high accuracy, are vulnerable to adversarial attacks. We proposed the DeltaBound attack: a novel, powerful attack in the hard-label setting with $\ell_2$ norm bounded perturbations. In this scenario, the attacker has only access to the top-1 predicted label of the model and can be therefore applied to real-world settings such as remote API. This is a complex problem since the attacker has very little information about the model. Consequently, most of the other techniques present in the literature require a massive amount of queries for attacking a single example. Oppositely, this work mainly focuses on the evaluation of attack's power in the low queries regime $\leq 1000$ queries) with $\ell_2$ norm in the hard-label settings. We find that the DeltaBound attack performs as well and sometimes better than current state-of-the-art attacks while remaining competitive across different kinds of models. Moreover, we evaluate our method against not only deep neural networks, but also non-deep learning models, such as Gradient Boosting Decision Trees and Multinomial Naive Bayes.

Lorenzo Rossi, Michael Aerni, Jie Zhang et al.

Sequence models, such as Large Language Models (LLMs) and autoregressive image generators, have a tendency to memorize and inadvertently leak sensitive information. While this tendency has critical legal implications, existing tools are insufficient to audit the resulting risks. We hypothesize that those tools' shortcomings are due to mismatched assumptions. Thus, we argue that effectively measuring privacy leakage in sequence models requires leveraging the correlations inherent in sequential generation. To illustrate this, we adapt a state-of-the-art membership inference attack to explicitly model within-sequence correlations, thereby demonstrating how a strong existing attack can be naturally extended to suit the structure of sequence models. Through a case study, we show that our adaptations consistently improve the effectiveness of memorization audits without introducing additional computational costs. Our work hence serves as an important stepping stone toward reliable memorization audits for large sequence models.

Lorenzo Rossi, Jacob Chakareski, Pascal Frossard et al.

Multiview video has recently emerged as a means to improve user experience in novel multimedia services. We propose a new stochastic model to characterize the traffic generated by a Multiview Video Coding (MVC) variable bit rate source. To this aim, we resort to a Poisson Hidden Markov Model (P-HMM), in which the first (hidden) layer represents the evolution of the video activity and the second layer represents the frame sizes of the multiple encoded views. We propose a method for estimating the model parameters in long MVC sequences. We then present extensive numerical simulations assessing the model's ability to produce traffic with realistic characteristics for a general class of MVC sequences. We then extend our framework to network applications where we show that our model is able to accurately describe the sender and receiver buffers behavior in MVC transmission. Finally, we derive a model of user behavior for interactive view selection, which, in conjunction with our traffic model, is able to accurately predict actual network load in interactive multiview services.