Fatima M. Ezzeddine

Fatima Ezzeddine, Luca Luceri, Omran Ayoub et al.

The detection of state-sponsored trolls operating in influence campaigns on social media is a critical and unsolved challenge for the research community, which has significant implications beyond the online realm. To address this challenge, we propose a new AI-based solution that identifies troll accounts solely through behavioral cues associated with their sequences of sharing activity, encompassing both their actions and the feedback they receive from others. Our approach does not incorporate any textual content shared and consists of two steps: First, we leverage an LSTM-based classifier to determine whether account sequences belong to a state-sponsored troll or an organic, legitimate user. Second, we employ the classified sequences to calculate a metric named the "Troll Score", quantifying the degree to which an account exhibits troll-like behavior. To assess the effectiveness of our method, we examine its performance in the context of the 2016 Russian interference campaign during the U.S. Presidential election. Our experiments yield compelling results, demonstrating that our approach can identify account sequences with an AUC close to 99% and accurately differentiate between Russian trolls and organic users with an AUC of 91%. Notably, our behavioral-based approach holds a significant advantage in the ever-evolving landscape, where textual and linguistic properties can be easily mimicked by Large Language Models (LLMs): In contrast to existing language-based techniques, it relies on more challenging-to-replicate behavioral cues, ensuring greater resilience in identifying influence campaigns, especially given the potential increase in the usage of LLMs for generating inauthentic content. Finally, we assessed the generalizability of our solution to various entities driving different information operations and found promising results that will guide future research.

Tomasz Szandala, Fatima Ezzeddine, Natalia Rusin et al.

Artificial Intelligence-generated content has become increasingly popular, yet its malicious use, particularly the deepfakes, poses a serious threat to public trust and discourse. While deepfake detection methods achieve high predictive performance, they often exhibit biases across demographic attributes such as ethnicity and gender. In this work, we tackle the challenge of fair deepfake detection, aiming to mitigate these biases while maintaining robust detection capabilities. To this end, we propose a novel post-processing approach, referred to as Fairness-Oriented Final Layer Input Prioritising (Fair-FLIP), that reweights a trained model's final-layer inputs to reduce subgroup disparities, prioritising those with low variability while demoting highly variable ones. Experimental results comparing Fair-FLIP to both the baseline (without fairness-oriented de-biasing) and state-of-the-art approaches show that Fair-FLIP can enhance fairness metrics by up to 30% while maintaining baseline accuracy, with only a negligible reduction of 0.25%. Code is available on Github: https://github.com/szandala/fair-deepfake-detection-toolbox

Fatima Ezzeddine, Omran Ayoub, Silvia Giordano

In recent years, there has been a notable increase in the deployment of machine learning (ML) models as services (MLaaS) across diverse production software applications. In parallel, explainable AI (XAI) continues to evolve, addressing the necessity for transparency and trustworthiness in ML models. XAI techniques aim to enhance the transparency of ML models by providing insights, in terms of model's explanations, into their decision-making process. Simultaneously, some MLaaS platforms now offer explanations alongside the ML prediction outputs. This setup has elevated concerns regarding vulnerabilities in MLaaS, particularly in relation to privacy leakage attacks such as model extraction attacks (MEA). This is due to the fact that explanations can unveil insights about the inner workings of the model which could be exploited by malicious users. In this work, we focus on investigating how model explanations, particularly counterfactual explanations (CFs), can be exploited for performing MEA within the MLaaS platform. We also delve into assessing the effectiveness of incorporating differential privacy (DP) as a mitigation strategy. To this end, we first propose a novel approach for MEA based on Knowledge Distillation (KD) to enhance the efficiency of extracting a substitute model of a target model exploiting CFs, without any knowledge about the training data distribution by the attacker. Then, we advise an approach for training CF generators incorporating DP to generate private CFs. We conduct thorough experimental evaluations on real-world datasets and demonstrate that our proposed KD-based MEA can yield a high-fidelity substitute model with a reduced number of queries with respect to baseline approaches. Furthermore, our findings reveal that including a privacy layer can allow mitigating the MEA. However, on the account of the quality of CFs, impacts the performance of the explanations.