Split Federated Learning (SFL) is an emerging paradigm for privacy-preserving distributed learning. However, it remains vulnerable to sophisticated data poisoning attacks targeting local features, labels, smashed data, and model weights. Existing defenses, primarily adapted from traditional Federated Learning (FL), are less effective under SFL due to limited access to complete model updates. This paper presents HealSplit, the first unified defense framework tailored for SFL, offering end-to-end detection and recovery against five sophisticated types of poisoning attacks. HealSplit comprises three key components: (1) a topology-aware detection module that constructs graphs over smashed data to identify poisoned samples via topological anomaly scoring (TAS); (2) a generative recovery pipeline that synthesizes semantically consistent substitutes for detected anomalies, validated by a consistency validation student; and (3) an adversarial multi-teacher distillation framework trains the student using semantic supervision from a Vanilla Teacher and anomaly-aware signals from an Anomaly-Influence Debiasing (AD) Teacher, guided by the alignment between topological and gradient-based interaction matrices. Extensive experiments on four benchmark datasets demonstrate that HealSplit consistently outperforms ten state-of-the-art defenses, achieving superior robustness and defense effectiveness across diverse attack scenarios.

The deep learning-based semantic communication system (DeepSC) is designed to improve the efficiency and accuracy of information transmission, by leveraging joint source-channel coding techniques to extract relevant semantic features. However, existing research on attack methods targeting DeepSC has primarily focused on text and image domains, leaving the speech domain largely unexplored. To this end, this paper proposes Iterative Semantic Gradient Update (ISGU), a novel approach for crafting physical layer adversarial attacks on DeepSC for speech transmission (DeepSC-ST). Specifically, we introduce a joint loss that combines the semantic similarity loss with Connectionist Temporal Classification loss to expedite the generation process of targeted attacks against the DeepSC-ST. In addition, we design an algorithm to generate adversarial examples, enhancing their imperceptibility by meticulously controlling the perturbation power added to the input speech. Extensive experiments indicate that ISGU is capable of rapidly generating highly covert adversarial examples, with a notably high attack success rate.

Coupled with the rise of Deep Learning, the wealth of data and enhanced computation capabilities of Internet of Things (IoT) components enable effective artificial intelligence (AI)-based models to be built. Beyond ground data sources, autonomous aerial vehicles (AAVs)-based service providers for data collection and AI model training, i.e., Drones-as-a-Service (DaaS), have become increasingly popular in recent years. However, the stringent regulations governing data privacy potentially impede data sharing across independently owned AAVs. To this end, we propose in this article a federated learning (FL)-based architecture that enables privacy-preserving collaborative machine learning across a federation of independent DaaS providers for the development of IoT applications. Specifically, this work introduces a novel incentive mechanism based on the hedonic coalition formation game to enhance the sustainable efficiency and stability of the FL system. By establishing tailored operational rules and functions, the proposed mechanism enables IoT sensing nodes to autonomously form optimal coalitions with AAVs, thereby ensuring robust collaboration. To deal with incentive mismatches and information asymmetry, we leverage the contract theory and propose a self-disclosure mechanism that guarantees truthful reporting of AAV capabilities while optimizing the global model owner’s profits. The performance-based AAV type is also defined to offer a practical measure for heterogeneous AAVs and serve as the foundation for fair and effective contract design. Simulation results validate the superiority of the proposed approach, demonstrating significant improvements in utility optimization and system stability compared to existing benchmarks.

Social networks are now the primary channels for sharing information about events, especially during emergencies. Identifying crucial information in these situations is intricate due to the absence of specific event labels and the prevalent label imbalance in social network data. Prior efforts in disaster event management have applied unsupervised domain adaptation (UDA) techniques. However, these efforts encounter difficulties when focusing on specific events and dealing with class-imbalance domain adaptation (CDA), often leading to modifications in data distribution. This article introduces a solution named class-imbalanced adversarial neural network (CADNN) to address UDA challenges in emergency events across diverse domains. First, an adversarial domain adaptation model is developed to achieve cross-domain feature representation. Second, unlike existing methods that rely on data augmentation, a cross-domain interpolation model is designed to generate aligned pairwise samples. This ensures that the generated samples preserve the original data distribution by preserving the sample centroid. Third, the sample interpolation is integrated with the adversarial domain adaptation model, optimizing the overall loss function to enhance generalization performance. Experimental results on four open-source datasets demonstrate that CADNN outperforms state-of-the-art models in recognizing emergency event information.

In mobile edge networks, federated learning (FL) has garnered substantial attention as a distributed machine learning framework with significant advantages for protecting user privacy. Due to the limited resources of wireless bandwidth, such FL-based applications are quite susceptible to Distributed Denial-of-Service (DDoS) attacks. Prior solutions either rely on centralized mechanisms that require complete information about all participants or are customized to specific systems. However, these solutions are either obsolete or ineffective given the new properties of FL. In this work, we first formulate a DDoS mitigation problem on bandwidth allocation for FL within mobile edge networks. Considering interactions between various network components and users, we propose anEvolutionaryGame andDouble-sidedAuction-based framework, termed EGDA, which consists of EG-based and DA-based mechanisms for user-bandwidth allocation (UBA) and server-bandwidth allocation (SBA), respectively. Specifically, to address DDoS attacks on UBA, we design an EG-based approach with minimum latency for FL under limited information. The proposed EG-based allocation algorithm is proven to be stable and achieve the evolutionary equilibrium. To mitigate DDoS attacks on SBA, we study an approach of DA with social welfare maximization while protecting the privacy of participants. Then, an iterative DA-based allocation algorithm is developed to be convergent and satisfy desirable economic properties. Extensive evaluation demonstrates that EGDA mitigates DDoS attacks effectively and efficiently.

Distributed Vehicle Sharing System (DVSS) leverages emerging technologies such as blockchain to create a secure, transparent, and efficient platform for sharing vehicles. In such a system, both efficient matching of users with available vehicles and optimal pricing mechanisms play crucial roles in maximizing system revenue. However, most existing schemes utilize user-to-vehicle (two-sided) matching and pricing, which are unrealistic for DVSS due to the lack of participation of service providers. To address this issue, we propose in this paper a novel Three-sided stable Matching with an optimal Pricing (TRIMP) scheme. First, to achieve maximum utilities for all three parties simultaneously, we formulate the optimal policy and pricing problem as a three-stage Stackelberg game and derive its equilibrium points accordingly. Second, relying on these solutions from the Stackelberg game, we construct a three-sided cyclic matching for DVSS. Third, as the existence of such a matching is NP-complete, we design a specific vehicle sharing algorithm to realize stable matching. Extensive experiments demonstrate the effectiveness of our TRIMP scheme, which optimizes the matching process and ensures efficient resource allocation, leading to a more stable and well-functioning decentralized vehicle sharing ecosystem.

Many vehicular applications, especially safety-related ones, rely on spatial-temporal messages periodically broadcast by vehicles. In the absence of a secure authentication scheme, invalid spatial-temporal messages may be sent out by malicious vehicles. Meanwhile, malicious applications may also collect a lot of personal information from spatial-temporal messages. Since inter-vehicle connections are often deployed in high-moving traffic, any authentication must be implemented in real-time. To meet all these properties, we propose a Fast and Anonymous Spatial-Temporal Trust (FastTrust) scheme for inter-vehicle connections. In contrast to most authentication protocols which rely on fixed infrastructures, FastTrust is mostly designed on hash chains and an entropy-based commitment, and is able to secure periodic spatial-temporal messages. FastTrust also protects vehicles’ privacy by deploying a pseudonym-varying scheduling mechanism to satisfy the anonymity and unlinkability requirements. Finally, in order to efficiently isolate malicious vehicles, a lightweight certificate management scheme is proposed for the limited bandwidth of vehicular networks. We provide analytical evaluations to show that our FastTrust achieves the security and privacy properties. Extensive validations are done to show that FastTrust can authenticate dozens of times faster than the existing signature algorithms, and isolate malicious vehicles at a low cost in terms of communication and computational resources.

User-Review Social Networks (URSNs) now become the targets of Sybil attacks, where fake reviews are posted by attackers to raise the reputation of listed services or products. Unlike previous fake accounts on Twitter or Weibo, Sybil attackers on URSNs are also genuine users in many cases, which presents a challenge for the existing fake review detection system. Hence, the main purpose of detecting Sybil attackers is to profile abnormal behaviors of users, and we propose a novel Sybil detection model named MUSH to extract long-term features of user behaviors on URSNs. First, aiming to measure the uncertainty of user behaviors, four entropy-based preference models are designed to quantify user preferences, including catering preferences, price preferences, word-of-mouth preferences, and rating preferences. Second, in order to extract temporal logic features, a novel multi-stimuli Hawkes process is introduced by combining external incentives and internal incentives, which can detect abnormal event sequences of posted reviews. This approach is quite different from previous solutions which mostly use direct/indirect graph models or user-related features for detection. Finally, by integrating entropy-based preferences with temporal logic features, a smart Sybil detection model is proposed based on a binary classification approach. Extensive experimental results indicate that MUSH can effectively detect Sybil attackers with a high detection accuracy. By comparing with other approaches, MUSH is quite suitable for adversarial network environments, which could provide better security services to social networks.

Mobile Social Networks have become one of the most convenient services for users to share information everywhere. This crowdsourced information is often meaningful and recommended to users, e.g., reviews on Yelp or high marks on Dianping, which poses the threat of Sybil attacks. To address the problem of Sybil attacks, previous solutions mostly use indirect/direct graph model or clickstream model to detect fake accounts. However, they are either dependent on strong connections or solely preserved by servers of social networks. In this paper, we propose a novel predictable approach by exploiting users' custom patterns to distinguish Sybil attackers from normal users for the application of recommendation in mobile social networks. First, we introduce the entropy of spatial-temporal features to profile the mobility traces of normal users, which is quite different from Sybil attackers. Second, we develop discriminative entropy-based features, i.e., users' preference features, to measure the uncertainty of users' behaviors. Third, we design a smart Sybil detection model based on a binary classification approach by combining our entropy-based features with traditional behavior-based features. Finally, we examine our model and carry out extensive experiments on a real-world dataset from Dianping. Our results have demonstrated that the model can significantly improve the detection accuracy of Sybil attacks.

The new coronavirus epidemic (COVID-19) has received widespread attention, causing the health crisis across the world. Massive information about the COVID-19 has emerged on social networks. However, not all information disseminated on social networks is true and reliable. In response to the COVID-19 pandemic, only real information is valuable to the authorities and the public. Therefore, it is an essential task to detect rumors of the COVID-19 on social networks. In this paper, we attempt to solve this problem by using an approach of machine learning on the platform of Weibo. First, we extract text characteristics, user-related features, interaction-based features, and emotion-based features from the spread messages of the COVID-19. Second, by combining these four types of features, we design an intelligent rumor detection model with the technique of ensemble learning. Finally, we conduct extensive experiments on the collected data from Weibo. Experimental results indicate that our model can significantly improve the accuracy of rumor detection, with an accuracy rate of 91% and an AUC value of 0.96.

Connected cars have received massive attention in Intelligent Transportation System. Many potential services, especially safety-related ones, rely on spatial-temporal messages periodically broadcast by cars. Without a secure authentication algorithm, malicious cars may send out invalid spatial-temporal messages and then deny creating them. Meanwhile, a lot of private information may be disclosed from these spatial-temporal messages. Since cars move on expressways at high speed, any authentication must be performed in real-time to prevent crashes. In this paper, we propose a Fast and Anonymous Spatial-Temporal Trust (FastTrust) mechanism to ensure these properties. In contrast to most authentication protocols which rely on fixed infrastructures, FastTrust is distributed and mostly designed on symmetric-key cryptography and an entropy-based commitment, and is able to fast authenticate spatial-temporal messages. FastTrust also ensures the anonymity and unlinkability of spatial-temporal messages by developing a pseudonym-varying scheduling scheme on cars. We provide both analytical and simulation evaluations to show that FastTrust achieves the security and privacy properties. FastTrust is low-cost in terms of communication and computational resources, authenticating 20 times faster than existing Elliptic Curve Digital Signature Algorithm.

Wireless sensor networks (WSNs) are increasingly used in smart cities which involve multiple city services having quality of service (QoS) requirements. When misbehaving devices exist, the performance of current delivery protocols degrades significantly. Nonetheless, the majority of existing schemes either ignore the faulty behaviors’ variability and time-variance in city environments or focus on homogeneous traffic for traditional data services (simple text messages) rather than city services (health care units, traffic monitors, and video surveillance). We consider the problem of fault-aware multiservice delivery, in which the network performs secure routing and rate control in terms of fault activity dynamic metric. To this end, we first design a distributed framework to estimate the fault activity information based on the effects of nondeterministic faulty behaviors and to incorporate these estimates into the service delivery. Then we present a fault activity geographic opportunistic routing (FAGOR) algorithm addressing a wide range of misbehaviors. We develop a leaky-hop model and design a fault activity rate-control algorithm for heterogeneous traffic to allocate resources, while guaranteeing utility fairness among multiple city services. Finally, we demonstrate the significant performance of our scheme in routing performance, effective utility, and utility fairness in the presence of misbehaving sensors through extensive simulations.

In vehicular networks, broadcast communications are critically important, as many safety-related applications rely on single-hop beacon messages broadcast to neighbor vehicles. However, it becomes a challenging problem to design a broadcast authentication scheme for secure vehicle-to-vehicle communications. Especially when a large number of beacons arrive in a short time, vehicles are vulnerable to computation-based Denial of Service (DoS) attacks that excessive signature verification exhausts their computational resources. In this paper, we propose an efficient broadcast authentication scheme called Prediction-Based Authentication (PBA) to not only defend against computation-based DoS attacks, but also resist packet losses caused by high mobility of vehicles. In contrast to most existing authentication schemes, our PBA is an efficient and lightweight scheme since it is primarily built on symmetric cryptography. To further reduce the verification delay for some emergency applications, PBA is designed to exploit the sender vehicle’s ability to predict future beacons in advance. In addition, to prevent memory-based DoS attacks, PBA only stores shortened re-keyed Message Authentication Codes (MACs) of signatures without decreasing security. We analyze the security of our scheme and simulate PBA under varying vehicular network scenarios. The results demonstrate that PBA fast verifies almost 99 percent messages with low storage cost not only in high-density traffic environments but also in lossy wireless environments.

Many location-based services require a mobile user to continuously prove his location. In absence of a secure mechanism, malicious users may lie about their locations to get these services. Mobility trace, a sequence of past mobility points, provides evidence for the user's locations. In this paper, we propose a Continuous Location Integrity and Provenance (CLIP) Scheme to provide authentication for mobility trace, and protect users' privacy. CLIP uses low-power inertial accelerometer sensor with a light-weight entropy-based commitment mechanism and is able to authenticate the user's mobility trace without any cost of trusted hardware. CLIP maintains the user's privacy, allowing the user to submit a portion of his mobility trace with which the commitment can be also verified. Wireless Access Points (APs) or colocated mobile devices are used to generate the location proofs. We also propose a light-weight spatial-temporal trust model to detect fake location proofs from collusion attacks. The prototype implementation on Android demonstrates that CLIP requires low computational and storage resources. Our extensive simulations show that the spatial-temporal trust model can achieve high (> 0.9) detection accuracy against collusion attacks.

Vehicular Ad Hoc Networks (VANETs) enable vehicle-to-vehicle communication to enhance road safety and improve driving experience. To secure periodic single-hop beacon messages for VANET applications, digital signature is one of the fundamental security approaches. However, it is vulnerable as excessive signatures would exhaust the computational resources of vehicles. In this paper, we propose a novel authentication mechanism VSPT, VANET authentication with Signatures and Prediction-based TESLA, which combines the advantages of both Elliptic Curve Digital Signature Algorithm (ECDSA) and Prediction-based TESLA. Although ECDSA is computationally expensive, it provides authentication and non-repudiation. Prediction-based TESLA enables fast and efficient verification by exploiting the sender's ability to predict its own future beacons. Both theoretical analysis and simulation results show that VSPT outperforms either the signature or TESLA in not only lossless situations but also lossy environments.