International audience

Machine unlearning aims to remove specific data from models to meet privacy regulations. While prior work has explored potential vulnerabilities in unlearning mechanisms, most assume adversaries with privileged access—an unrealistic premise in real-world Machine Learning-as-a-Service (MLaaS) settings. This raises a fundamental question: Can unlearning be exploited under restricted access constraints? We answer this affirmatively by proposing Class-Aligned Unlearning Attack (CAUA), a novel attack framework tailored to realistic deployment settings. CAUA uses target-class examples and public out-of-distribution data to generate semantically aligned inputs that integrate seamlessly into training. These inputs maintain model performance during training but, when unlearned, induce localized representation collapse and significant shifts in decision boundaries. We comprehensively evaluate CAUA across multiple datasets and unlearning paradigms. Notably, unlearning just 0.2% of training data causes up to a 73.4% drop in target-class F1 and a 60.4% drop in overall accuracy, revealing a previously overlooked vulnerability. Our work sheds new light on the risks of machine unlearning and lays a foundation for building more robust defenses.

The rapid development of network technology has allowed numerous vehicular applications to be deployed in vehicles, thereby enriching the driving experience of users. However, the openness of vehicular networks enables attackers to launch network attacks on vehicles through network ports, leading to the destruction of vehicular networks. To develop an intrusion detection system suitable for distributed vehicular networks, researchers have utilized federated learning to train detection models. Nevertheless, most federated learning-based vehicular intrusion detection systems seldom consider rapidly updating the detection model and fail to detect unknown attacks effectively. In this study, we propose a federated learning-based vehicular intrusion detection system that fully considers the traffic characteristics of multiple network regions and selects representative clients to participate in model aggregation, thereby accelerating the convergence of the global model. Furthermore, to enhance the robustness of the detection system, we utilize extreme value theory and multilayer activation vectors to construct an unknown attack discriminator that can determine whether a network flow is an unknown attack. Comprehensive experiments on three open datasets demonstrate that the proposed intrusion detection system can quickly update and effectively identify known/unknown attacks in open vehicular networks

Providing high-quality network services for vehicles is a challenge because of the fast-moving character of the vehicles. To address the shortcomings of traditional centralized and distributed mobility management schemes, such as triangular routing and poor scalability, many researchers use software-defined networking (SDN) to build mobility management schemes. However, most schemes rarely consider how to select the optimal base station for high-speed mobile vehicles in a dense network environment. Only using the received signal strength to select the base station tends to cause a ping-pong effect. Moreover, due to the high mobility of vehicles, the routing updates between vehicles and communication nodes will frequently occur, resulting in the significant consumption of network resources. In this article, we propose a mobility management scheme MM-SDVN based on SDN for vehicles. MM-SDVN uses deep Q-network to construct the optimal base station selection model, designs a multipath prefix matching algorithm to reduce the cost of route update, and realizes the seamless handover of vehicles in SDN intradomain and interdomain scenarios. The comprehensive experimental results show that MM-SDVN greatly improves the network service quality and handover performance of the vehicle. Compared to the other schemes, MM-SDVN improved vehicle throughput by 6.14%, 8.85%, and 10.34%, respectively.

In vehicular networks, caching content in fog nodes is a widely accepted and favorable way to quickly respond to massive vehicle requests, reduce content retrieval delay and improve service quality. However, to implement such caching mode, it is critical to ensure efficient security and privacy protection when vehicles access the cached content in fog nodes. In this paper, aiming at the security issue, a novel lightweight cryptography-based access control scheme for software defined vehicular networks (SDVN) is proposed, by using TESLA broadcast authentication protocol and Pederson commitment. The scheme realizes direct and efficient authentication between vehicles and fog nodes while limiting only legitimate vehicles can get request responses, and avoids limitations or deficiencies in existing access control schemes. Moreover, considering the limited cache space of the fog node, by utilizing the flexibility of the SDN paradigm, a cooperative cache update mechanism is provided. The security verification with ProVerif and detailed security analyses prove that the scheme can meet the security requirements in SDVN. And compared with the related works, our scheme achieves better performance in terms of computation and communication costs.