Mobile Edge CrowdSensing (MECS) based on Federated Learning (FL) has attracted widespread attention as an intelligent data collection and processing approach. FL trains the global model through aggregating local models of participants without requiring the exchange of raw data. However, directly sharing local models is vulnerable to backdoor attacks launched by adversaries. What's worse, malicious server may collude with participants to manipulate the parameter updating of models and even compromise the accuracy of whole system. To address these challenges, this paper proposes a Privacy-Enhanced Federated Learning (PEFL) framework for MECS with the aim of resisting both backdoor and collusion attacks. In PEFL, a Backdoor Resistant Privacy-Enhanced Aggregation (BRPEA) mechanism with the Differential Privacy-Enhanced Exponential (DPEE) method is developed to perturb local models of participants. Clustering and clipping techniques are also designed in BRPEA to effectively distinguish backdoor models from the benign local models, which eliminate the influence of local models deviation and optimize the noise introduced by differential privacy. Furthermore, a Collusion Resistant Privacy-Preserving Aggregation (CRPEA) mechanism is studied. CRPEA can avoid the collusion between servers and participants and prevent the privacy of local models from being leaked. The theoretical analysis proves the security of proposed PEFL framework and the simulation experiments demonstrate that PEFL can not only ensure the aggregation accuracy of encrypted models but provide robustness against both backdoor and collusion attacks.

Mobile Crowd-Sensing (MCS) has emerged as a significant approach in various domains for collecting and disseminating sensing data. However, it is a challenging issue to select appropriate participants for a sensing task. This paper introduces a Quality Awareness Incentive Mechanism based on Taxation and Data Aggregation (QIM-TDA), which can choose the reliable participants without losing the platform utility and the data quality. Firstly, we define a reputation model to measure the reliability of participant and select more reliable participants based on their reputation in order to maximize platform utility. Secondly, a truth discovery algorithm is proposed to aggregate the sensing data and ensure the data quality of MCS. Finally, a normalized taxation mechanism is discussed in order to prevent the excessive accumulation of reputation for participant and further enhance the data quality. The simulation results prove that QIM-TDA can significantly improve the data quality and task completion rate compared to some typical mechanisms.

Data aggregation (DA) plays an important role in the context of Internet of Things (IoT). Although some favorable solutions have been proposed to improve the performances of DA, the complex collusion attacks are often ignored and may produce more serious negative impact on aggregation accuracy. In this article, we design a novel dynamic robust iterative filtering (DRIF) mechanism to enhance the quality of service of IoT applications and improve the vulnerability of DA to the collusion attack. First, the initial reputations based on the maximum likelihood estimation are assigned to sensor nodes in order to resist the collusion attack. Second, the sensor nodes obtain the aggregation result through iterative filtering so as to ensure the accuracy of DA. Especially, a weight updating scheme is proposed to eliminate the negative effect of the accidental anomaly or collusion nodes. Finally, the simulation study indicates that the proposed DRIF mechanism is effective and it can achieve a higher accuracy in the presence of complex dynamic collusion attacks.

Visual Transformers (ViTs) and Convolutional Neural Networks (CNNs) are the two primary backbone structures extensively used in various vision tasks. Generating transferable adversarial examples for ViTs is difficult due to ViTs’ superior robustness, while transferring adversarial examples across ViTs and CNNs is even harder, since their structures and mechanisms for processing images are fundamentally distinct. In this work, we propose a novel attack method named Momentum Integrated Gradients (MIG), which not only attacks ViTs with high success rate, but also exhibits impressive transferability across ViTs and CNNs. Specifically, we use integrated gradients rather than gradients to steer the generation of adversarial perturbations, inspired by the observation that integrated gradients of images demonstrate higher similarity across models in comparison to regular gradients. Then we acquire the accumulated gradients by combining the integrated gradients from previous iterations with the current ones in a momentum manner and use their sign to modify the perturbations iteratively. We conduct extensive experiments to demonstrate that adversarial examples obtained using MIG show stronger transferability, resulting in significant improvements over state-of-the-art methods for both CNN and ViT models.