Convolutional Neural Networks (CNNs) that have excelled in diverse computer vision tasks are vulnerable to backdoor attacks, enabling attacker-controlled predictions via specific triggers. Restricted to spatial domains, recent research exploits perceptual traits by embedding triggers in the frequency domain, yielding pixel-level indistinguishable perturbations. In black-box settings, restricted access to model and training process necessitates advanced trigger designs. Current frequency-based attacks manipulate magnitude spectra, introducing discrepancies between clean and poisoned data, though vulnerable to common image processing operations like compression and filtering. In this paper, we propose a robust low-frequency backdoor attack (LFBA) in black-box setup that minimally perturbs spectrum components and maintains the perceptual similarity in spatial space simultaneously. Our methodology capitalizes on the insight that optimal triggers can be located in low-frequency regions to maximize attack effectiveness, robustness against image transformation operations, and stealthiness in dual space. To effectively explore the discrete frequency space, we utilize simulated annealing (SA), a form of evolutionary algorithm, to optimize the properties of trigger including the frequency bands to be manipulated and the perturbation of each band under restricted attack scenario. Extensive experiments on both CNNs and Vision Transformers (ViT) confirm the effectiveness and robustness of LFBA against image processing operations and state-of-the-art backdoor defenses. Furthermore, LFBA exhibits inherent stealthiness in both spatial and frequency spaces, making it resistant to human and frequency inspection.

Vertical Federated Learning (VFL) enables multiple clients to collaboratively train a global model over vertically partitioned data without leaking private local information. Tree-based models, like XGBoost and LightGBM, have been widely used in VFL to enhance the interpretation and efficiency of training. However, there is a fundamental lack of research on how to conduct VFL securely over distributed labels. This work is the first to fill this gap by designing a novel protocol, called FEVERLESS, based on XGBoost. FEVERLESS leverages secure aggregation via information masking technique and global differential privacy provided by a fairly and randomly selected noise leader to prevent private information from being leaked in the training process. Furthermore, it provides label and data privacy against honest-but-curious adversaries even in the case of collusion of$n - 2$out of n clients. We present a comprehensive security and efficiency analysis for our design, and the empirical results from our experiments demonstrate that FEVERLESS is fast and secure. In particular, it outperforms the solution based on additive homomorphic encryption in runtime cost and provides better accuracy than the local differential privacy approach.

Recent studies have revealed the vulnerability of graph convolutional networks (GCNs) to edge-perturbing attacks, such as maliciously inserting or deleting graph edges. However, theoretical proof of such vulnerability remains a big challenge, and effective defense schemes are still open issues. In this article, we first generalize the formulation of edge-perturbing attacks and strictly prove the vulnerability of GCNs to such attacks in node classification tasks. Following this, an anonymous GCN, named AN-GCN, is proposed to defend against edge-perturbing attacks. In particular, we present a node localization theorem to demonstrate how GCNs locate nodes during their training phase. In addition, we design a staggered Gaussian noise-based node position generator and a spectral graph convolution-based discriminator (in detecting the generated node positions). Furthermore, we provide an optimization method for the designed generator and discriminator. It is demonstrated that the AN-GCN is secure against edge-perturbing attacks in node classification tasks, as AN-GCN is developed to classify nodes without the edge information (making it impossible for attackers to perturb edges anymore). Extensive evaluations verify the effectiveness of the general edge-perturbing attack (G-EPA) model in manipulating the classification results of the target nodes. More importantly, the proposed AN-GCN can achieve 82.7% in node classification accuracy without the edge-reading permission, which outperforms the state-of-the-art GCN.

Synthetic data generation plays a crucial role in many areas where data is scarce and privacy/confidentiality is a significant concern. Generative Adversarial Networks (GANs), arguably one of the most widely used data synthesis techniques, allow for the training of a model (i.e., generator) that can generate real-looking data by playing a min-max game with a discriminator model. When multiple organizations are reluctant to share their sensitive data, GANs models can be trained in a federated manner, commonly with the use of differential privacy (DP). In order to achieve a reasonable level of model utility, DP trades privacy exhibiting vulnerability to various attacks (e.g., membership inference attack). In this paper, we propose a hybrid solution, PP-FedGAN, to the asynchronous federated, privacy-preserving training of GANs models by combining the CKKS homomorphic encryption (HE) scheme with differential privacy. The addition of HE results in around 10 seconds of overhead on the client side per round and 115 seconds on the entire training procedure. We also analyze the security of PP-FedGAN under the honest-but-curious security model. Where stronger security guarantees are required, our proposal presents a better alternative to solutions that only employ DP.

Graph Neural Networks (GNNs) are a class of deep learning-based methods for processing graph domain information. GNNs have recently become a widely used graph analysis method due to their superior ability to learn representations for complex graph data. Due to privacy concerns and regulation restrictions, centralized GNNs can be difficult to apply to data-sensitive scenarios. Federated learning (FL) is an emerging technology developed for privacy-preserving settings when several parties need to train a shared global model collaboratively. Although several research works have applied FL to train GNNs (Federated GNNs), there is no research on their robustness to backdoor attacks.