Efficient and privacy-aware multi-party classification protocol for human activity recognition

https://doi.org/10.1016/j.jnca.2017.09.005Get rights and content

Abstract

Human activity recognition (HAR) is an important research field that relies on sensing technologies to enable many context-aware applications. Nevertheless, tracking personal signs to enable such applications has given rise to serious privacy issues, especially when using external activity recognition services. In this paper, we propose (Π-Knn): a privacy-preserving version of the K Nearest Neighbors (k-NN) classifier that is mainly built on (Π-CSP+): a novel cryptography-free private similarity evaluation protocol. As a sample application, we consider a medical monitoring system enhanced with a HAR process based on our privacy preserving classifier. The integration of the privacy preserving HAR aims to improve the accuracy of the clinical decision support. We conduct a standard security analysis to prove that our protocols provide a complete privacy protection against malicious adversaries. We perform a comparative performance evaluation through several experiments while using real HAR system parameters. Experimental evaluations show that our protocol (Π-CSP+) incurs a low increasing overhead (37% in Online classification and 50% in Offline classification) compared to PCSC, a representative state-of-the art protocol, which incurs 3600% and 4800% in online and offline classification respectively. Besides, Π-CSP+ provides a stable and efficient response time (W=0.0x ms) for both short and long duration activities while serving up to 1000 clients. Comparative results confirm the computational efficiency of our protocol against a competitive state-of-the-art protocol.

Introduction

Data mining methods are gaining an increasing attention because of the wide proliferation of knowledge-based applications. Analyzing data from wireless and sensor networks has enabled developing new services, such as Human Activity Recognition (HAR). HAR consists of tracking environmental and personal sensed signs, then, analyzing them to provide accurate information about persons’ daily activities. Nevertheless, the collection and analysis of personal private data, such as GPS location, raises concerns about users’ privacy, especially when the analysis is performed through external service providers. External recognition aims to reduce the cost of computation and storage accrued by client devices. Additionally, it aims ensuring a high accuracy level in recognition results, which are built upon big data stores of activity patterns.

To face such a concern, several Privacy-Preserving Data Mining (PPDM) methods have been proposed. These include classification, clustering and other data mining tasks (Sachan et al., 2013). PPDM methods protect the privacy by changing or deleting sensitive data before analysis (Xu and Yi, 2011). This approach is based on a trade-off between accuracy and privacy (Aldeen et al., 2015). Other approaches employed cryptographic techniques to provide a high privacy protection level, but, they are computationally very expensive (Lu et al., 2014).

From another side, privacy-preserving HAR may provide useful information that enhances context-aware aspects in several applications, like e-healthcare monitoring systems. CodeBlue (Malan et al., 2004), AlarmNet (Wood et al., 2006) and some other popular medical monitoring systems (Chakravorty, 2006, van Halteren et al., 2004) have been proposed and focused on addressing power, security and computational resource constraints (Kumar and Lee, 2011). Yet, they have some shortages in tracking information about patients’ physical activities. Such information is useful to avoid wrong diagnosis and treatment when vital sensed signs are jammed, errored or modified. To shed light on these shortages, studies on information needed by clinicians show that in about 81% of ambulatory diagnosis, physicians are missing critical information (Musen et al., 2014). Other studies report that about 18% of medical errors may be due to insufficient availability of patient information (Leape, 1994). Thus, acquiring a complete picture of patient state will reduce medical errors and may drive for a broad adoption of e-healthcare monitoring systems for the clinical decision support (CDS) task.

In this paper, we propose a novel privacy-preserving k-NN classification version, which aims to address privacy and efficiency concerns when using external services for human activity recognition. As an application, we propose a framework that combines the human activity recognition (HAR) process with the clinical decision support (CDS) process. This may enhance accuracy in medical decision while protecting patients’ privacy.

We summarize the contributions of this work in the following items

  • We build a novel privacy-preserving version of k-NN, named (Π-Knn), and we use it for the classification task, which is applied according to external activity patterns.

  • We propose (Π-CSP+), a novel privacy-preserving and efficient cosine similarity protocol, which is the main core of (Π-Knn). It aims to securely assess similarity between HAR sensed data and external activity patterns. Π-CSP+ is based on simple arithmetic operations to avoid computation overheads induced by cryptographic techniques.

  • As an application of the HAR system, we propose SimilCare, a novel medical monitoring framework that embeds information about patients’ activities within a clinical knowledge database while using our proposed Π-Knn protocol. SimilCare aims to cover shortage of existing healthcare monitoring systems in tracking information about patients’ activities, while ensuring their privacy.

  • We present a security analysis of our proposed protocols (Π-CSP+ and Π-Knn) using a standard security proof (Canetti, 2000), which has revealed a complete privacy protection. In addition, we perform simulations through different experiments while using real HAR system parameters. The computation performances are highly efficient compared to the most efficient protocol found in the literature (Lu et al., 2014).

The remainder of this paper is organized as follows. In Section 2, we provide a literature survey of related works and we discuss them. Section 3 presents preliminaries and building blocks used for designing our protocols. Next, we devote Section 4 to present our privacy-preserving protocols, besides their integration in the proposed SimilCare framework. Then, we evaluate the privacy protection and the performance level in Section 5 and Section 6 respectively. We end-up this work with our final conclusions in Section 7.

Section snippets

Related work

Several existing HAR systems have not considered protecting users’ privacy during the recognition and classification phase. In this section, we review recent works in HAR field. Besides, we give a review on privacy-preserving k-NN classification, and privacy-preserving similarity evaluation, which is the main privacy-related computation within k-NN protocol.

Human activity recognition (HAR)

Human activity recognition (HAR) is the field that aims to provide accurate information on people's activities. The general structure of a HAR system involves three main phases, as shown in Fig. 1.

  • In the data collection phase, the sensors’ raw data are communicated to the data collection node. Sensors are attached to different locations on the body or placed in the environment. The raw data are sampled in a multivariate time series (sij) depending on sensors frequencies, where j corresponds to

Π-Knn: A privacy-preserving and efficient k-nn classification protocol for human activity recognition

In this section, we present an efficient and privacy-preserving k-NN algorithm called Π-Knn. We build this protocol on a privacy-preserving Cosine Similarity Protocol that we call Π-CSP+. Next, we integrate these protocols in SimilCare, a novel proposed medical monitoring framework.

Security analysis

In this section, we provide a security analysis of our proposal according to the real/ideal simulation paradigm (Canetti, 2000, Lindell and Pinkas, 2009). We stress that such a proof provides very strong security guarantees (Lindell and Pinkas, 2009).

Note 3

Notice for clarification that real/ideal simulation given in this section has no relation with simulation made for the performance evaluation in the next section.

Performance analysis

  • Computation cost. In this section, we evaluate the computation performance of Π-CSP+ (Algorithm 2), which is the main core of the proposed Π-Knn protocol (Algorithm 3). This evaluation aims to analyze the effect of adding our privacy-preserving measurements through Π-CSP+ on the computational performance of the k-NN classifier. To do so, we consider a global context where a SimilCare HAR service denoted SC monitors a patient all day long. SC extracts (v) vectors of (n) features from the patient

Conclusion

In this paper, we have proposed a secure k-NN classification protocol named (Π-Knn), designed for Human Activity Recognition (HAR). We have built this protocol on a novel efficient and privacy-preserving cosine similarity protocol named (Π-CSP+). As an application, we have integrated our proposed privacy-preserving HAR classifier in SimilCare, a novel medical monitoring framework, to support the medical decision by securely providing information about patients’ activities. Through security

References (43)

  • Y.A.A.S. Aldeen et al.

    A comprehensive review on privacy preserving data mining

    SpringerPlus

    (2015)
  • K. Altun et al.

    Human Activity Recognition Using Inertial/Magnetic Sensor Units

    (2010)
  • O. Banos et al.

    Window size impact in human activity recognition

    Sensors

    (2014)
  • R. Canetti

    Security and composition of multiparty cryptographic protocols

    J. Cryptol.

    (2000)
  • Chakravorty, R., 2006. A programmable service architecture for mobile medical care. In: Proceedings of the Fourth...
  • J. Cheng et al.

    Active Capacitive Sensing: Exploring a New Wearable Sensing Modality for Activity Recognition

    (2010)
  • D. De et al.

    Multimodal wearable sensing for fine-grained activity recognition in healthcare

    IEEE Internet Comput.

    (2015)
  • Du, W., Atallah, M., Privacy-preserving cooperative statistical analysis. In: Proceedings of the 17th Annual Computer...
  • Evani, A., Sreenivasan, B., Sudesh, J., Prakash, M., Bapat, J., 2013. Activity recognition using wearable sensors for...
  • E. Gelenbe et al.
    (1987)
  • B. Goethals et al.

    On Private Scalar Product Computation for Privacy-Preserving Data Mining

    (2005)
  • Hou, J.C., Wang, Q., AlShebli, B.K., Ball, L., Birge, S., Caccamo, M., Cheah, C.-F., Gilbert, E., Gunter, C.A., Gunter,...
  • H. Huang et al.

    Secure two-party distance computation protocol based on privacy homomorphism and scalar product in wireless sensor networks

    Tsinghua Sci. Technol.

    (2016)
  • Huynh, T., Schiele, B., 2005. Analyzing features for activity recognition. In: Proceedings of the 2005 Joint Conference...
  • Jiang, S., Cao, Y., Iyengar, S., Kuryloski, P., Jafari, R., Xue, Y., Bajcsy, R., Wicker, S., 2008. Carenet: An...
  • Jiang, W., Murugesan, M., Clifton, C., Si, L., 2008. Similar document detection with limited information...
  • Kikuchi, H., Nagai, K., Ogata, W.,Nishigaki, M., 2008. Privacy-preserving similarity evaluation and application to...
  • P. Kumar et al.

    Security issues in healthcare applications using wireless medical sensor networks: a survey

    Sensors

    (2011)
  • O.D. Lara et al.

    A survey on human activity recognition using wearable sensors

    IEEE Commun. Surv. Tutor.

    (2013)
  • Lau, S.L., König, I., David, K., Parandian, B., Carius-Düssel, C., Schultz, M., 2010. Supporting patient monitoring...
  • Leontiadis, I., Önen, M., Molva, R., Chorley, M.J., Colombo, G.B., 2013. Privacy preserving similarity detection for...
  • Cited by (0)

    View full text