6 September 2021 in IEEE Internet of Things Journal
Both voice communication and automatic speech verification (ASV) over smart devices are vulnerable to the voice impersonation (VI) attack, which is often launched via imitating a target’s voice characteristics to deceive human auditory sense or fool the ASV system. Researchers have designed a number of defense schemes yet without the consideration of universality due to the lack of comprehensive data sets. In this article, we propose a universal defense scheme based on the VI data set collected from a famous TV show named “The Sound.” First, we deliver a thorough study on the VI attacks in both auditory and ASV systems to verify the collected simulated voice could spoof the auditory and the ASV system with a notable probability. Second, we propose a quasi-Gaussian distribution (QGD)-based defense scheme with the discovery about specific voice characteristics that are distinct between attackers and targets. Finally, we conduct extensive experimental results on our collected VI data set as well as the auxiliary ASVspoof2017 data set, to indicate the proposed QGD scheme outperforms the state-of-the-art schemes: backpropagation neural network, support vector machine, and Gaussian mixture model, in terms of accuracy.
15 May 2024 in IEEE Transactions on Mobile Computing
Numerous mobile devices are equipped with voice assistants to facilitate contactless user-device interaction. However, the widespread availability of voice assistants also raises security and privacy concerns, as they can be maliciously triggered to perform voice eavesdropping. Although diverse attacks have been taken to manipulate voice assistants for eavesdropping, they exhibit deficiencies of limited attack scopes and conspicuous attack behaviors because they target specific voice assistants or require extra voice commands to activate them. To manipulate arbitrary voice assistants for covert eavesdropping attack, we conduct a comprehensive analysis of voice assistant implementation in the Android system and refine a universal workflow. Through meticulous analysis and experimental verification, we uncover an inherent vulnerability that in voice assistants across device types that can be awakened by an artificial faking Intent. Building on this significant discovery, we propose an attack termed VoiceEar . It leverages a malicious event generation file and a first-in-first-out Intent generation algorithm to trigger voice assistants within the normal workflow for eavesdropping, without voice commands. Finally, we deploy the VoiceEar attacks on 25 mainstream mobile devices, and invite 95 volunteers for eavesdropping activity perception testing. The results unequivocally demonstrate the seamless execution of VoiceEar attacks, with neither users nor devices awareness.
8 July 2022 in IEEE Transactions on Network Science and Engineering
Anomalous traffic detection is a vital task in advanced Internet supervision and maintenance. To detect anomalies accurately, various data representations, such as vectors, matrices, and tensors, have been adopted to model traffic data. Among them, tensor-based methods outperform others due to their capability of capturing comprehensive correlations between complex network traffic. However, existing tensor-based algorithms remain certain shortcomings, such as working offline, cannot timely detect traffic anomalies, and high computation costs. To conquer the aforementioned deficiencies, we propose a novel sequence tensor recovery (STR) algorithm in this paper, which utilizes the results of historical tensor decomposition to achieve quick and accurate anomaly detection with low consumption when traffic data series arrive. Furthermore, we propose a dynamic sequence tensor recovery (DSTR) algorithm to improve anomaly detection accuracy by better capturing the variation over time of the comprehensive correlation of traffic data hidden in the tensor structure. The experimental results on two real traffic traces, Abilene and G $\grave{E}$ ANT, indicate the proposed STR and DSTR algorithms are superior to the state-of-the-art algorithms in terms of accuracy and computation cost.
18 November 2025 in IEEE Transactions on Mobile Computing
The mobile health (mHealth) system, leveraging mobile edge computing, can monitor health status and provide diagnosis. However, due to the privacy of medical data and the resource limitations of mobile devices, patients are unable to access diagnostic services provided by untrusted servers in real-time. Existing schemes present significant challenges in private heterogeneous data aggregation, model training and inference in the presence of malicious participants, and expensive resource consumption. To address these issues, in this paper, we propose a non-interactive privacy-preserving system with the naive Bayesian model, i.e., $\mathsf {RobustHealth}$, for heterogeneous mHealth diagnosis. Specifically, we extract homogeneous features from heterogeneous datasets to enable efficient encrypted aggregation. We propose a novel private model training algorithm with enhanced security to against collusion-then-differential attacks. We develop a novel non-interactive private model inference algorithm using minimal lightweight cryptographic primitives, designed for patients under unstable network environments. We provide formal security proofs for our system using the Universal Composable (UC) framework. To validate the performance of $\mathsf {RobustHealth}$, we conduct extensive experiments on real-world heterogeneous datasets, and compared with related works. The results demonstrate a $\mathbf {4.37\%}$ improvement in model accuracy, along with significant reductions in computational and communication overheads of $\mathbf {21.18\times }$ and $\mathbf {4.24\times }$, respectively.
10 December 2025 in IEEE Transactions on Dependable and Secure Computing
Deep learning technologies have seen widespread adoption in WiFi-based wireless sensing systems. However, they are inherently vulnerable to adversarial perturbation attacks, which has received little attention within the WiFi sensing community. To more comprehensively understand the potential threats posed by perturbation attacks, we present a novel attack method, named WiIntruder, distinguishing itself with universality, robustness, and stealthiness. This paper intends to provide a catalyst that promotes the assessment of security in existing WiFi-based sensing systems. We achieve the three aforementioned salient features in WiIntruder through the following three steps: (1) Maximizing transferability by differentiating user-state-specific feature spaces across sensing models, thereby enabling a universal perturbation attack vector applicable to a wide range of applications; (2) Mitigating the impact of perturbation signal distortion by optimizing key factors of device synchronization and wireless propagation through a heuristic particle swarm algorithm; and (3) Enhancing the diversity and stealthiness of attack patterns by randomly switching among perturbation surrogates generated by a generative adversarial network. Experimental results confirm the threat posed by WiIntruder to four common WiFi-based services, with the average accuracy decrease by 72.9% under black-box attack scenarios.
1 January 2025 in IEEE Transactions on Consumer Electronics
WiFi-based indoor positioning has emerged as a crucial technology for enabling smart consumer electronic applications, particularly in large-scale buildings. The construction of WiFi fingerprint databases using received signal strength (RSS) is foundational due to its widespread deployment. However, achieving high positioning accuracy typically requires labor-intensive and time-consuming site surveys. While recent crowdsourcing methods have facilitated the collection of numerous RSS samples, these samples frequently lack labels and reliability in multi-scale building environments.In this paper, we design a novel semi-supervised and multi-task mean-teacher model (MTMT-DNN) to annotate crowdsourcing unlabeled multi-scale fingerprint samples. This method enables the construction of a comprehensive fingerprint database without requiring intensive manual effort or compromising positioning accuracy. Our key idea is to first develop a multi-task Deep Neural Network (MT-DNN) for simultaneously annotating building, floor, and intra-floor coordinate labels by leveraging their complementary information. Then we employ the mean-teacher semi-supervised learning to leverage additional unlabeled fingerprint data for further improving the annotating performance and reducing intensive manual effort. Finally, we train the MTMT-DNN model by developing two multi-task loss functions and ensuring consistency between them, thereby enhancing the reliability of the annotated crowdsourced fingerprints. We conducted real-world experiments in a 20,000 m2 site encompassing three multi-story buildings. The results demonstrate that our proposed method significantly reduces the workload of manually collecting labeled fingerprint samples. With only 20% of labeled fingerprints collected, we achieve 99% average annotation accuracy for building and floor labels and an average coordinates annotation error within 4 m.