Membership Inference Attacks (MIAs) are currently a dominant approach for evaluating privacy in machine learning applications. Despite their significance in identifying records belonging to the training dataset, several concerns remain unexplored, particularly with regard to tabular data. In this paper, first, we provide an extensive review and analysis of MIAs considering two main learning paradigms: centralized and federated learning. We extend and refine the taxonomy for both. Second, we demonstrate the efficacy of MIAs in tabular data using several attack strategies, also including defenses. Furthermore, in a federated learning scenario, we consider the threat posed by an outsider adversary, which is often neglected. Third, we demonstrate the high vulnerability of single-outs (records with a unique signature) to MIAs. Lastly, we explore how MIAs transfer across model architectures. Our results point towards a general poor performance of these attacks in tabular data which contrasts with previous state-of-the-art. Notably, even attacks with limited attack performance can still successfully expose a large portion of single-outs. Moreover, our findings suggest that using different surrogate models makes MIAs more effective.

Federated Learning (FL) is a distributed machine learning approach that promises privacy by keeping the data on the device. However, gradient reconstruction and membership-inference attacks show that model updates still leak information. Fully Homomorphic Encryption (FHE) can address those privacy concerns but it suffers from ciphertext expansion and requires prohibitive overhead on resource-constrained devices. We propose the first Hybrid Homomorphic Encryption (HHE) framework for FL that pairs the PASTA symmetric cipher with the BFV FHE scheme. Clients encrypt local model updates with PASTA and send both the lightweight ciphertexts and the PASTA key (itself BFV-encrypted) to the server, which performs a homomorphic evaluation of the decryption circuit of PASTA and aggregates the resulting BFV ciphertexts. A prototype implementation, developed on top of the Flower FL framework, shows that on independently and identically distributed MNIST dataset with 12 clients and 10 training rounds, the proposed HHE system achieves 97.6% accuracy, just 1.3% below plaintext, while reducing client upload bandwidth by over 2,000x and cutting client runtime by 30% compared to a system based solely on the BFV FHE scheme. However, server computational cost increases by roughly 15621x for each client participating in the training phase, a challenge to be addressed in future work.