Preserving worker privacy in crowdsourcing

Abstract

This paper proposes a crowdsourcing quality control method with worker-privacy preservation. Crowdsourcing allows us to outsource tasks to a number of workers. The results of tasks obtained in crowdsourcing are often low-quality due to the difference in the degree of skill. Therefore, we need quality control methods to estimate reliable results from low-quality results. In this paper, we point out privacy problems of workers in crowdsourcing. Personal information of workers can be inferred from the results provided by each worker. To formulate and to address the privacy problems, we define a worker-private quality control problem, a variation of the quality control problem that preserves privacy of workers. We propose a worker-private latent class protocol where a requester can estimate the true results with worker privacy preserved. The key ideas are decentralization of computation and introduction of secure computation. We theoretically guarantee the security of the proposed protocol and experimentally examine the computational efficiency and accuracy.

Appendix 1: Extensions to multi-class and real-valued labels

We introduce the detailed update rules of modified LC methods to deal with multi-class and real-valued labels, and then we explain how to extend the inference algorithms to preserve worker privacy.

1.1 Appendix 1.1: Multi-class labels

The LC method was originally proposed for multi-class labels by Dawid and Skene (1979). Let us assume a task to give a $K$-class label ($K\ge 2$). For each $i\in {\mathcal I}$ and $j\in {\mathcal {J}}$, a crowd label $y_{i,j}\in \{0,\dots ,K-1\}(=:{\mathcal K})$ is generated by the multinomial distribution

$$\begin{aligned} \pi _{jkl} = \Pr [y_{i,j} = k \mid y_{i}=l, \Pi _{j}], \end{aligned}$$

where $\sum _{k\in {\mathcal K}} \pi _{jkl} = 1$ holds for all $l\in {\mathcal K}$, and we denote $\Pi _{j} = \{\pi _{jkl} \mid k,l\in {\mathcal K}\}$. Also, for each $i\in {\mathcal I}$, the true label $y_i\in {\mathcal K}$ is generated by

$$\begin{aligned} p_l = \Pr [y_{i} = l], \end{aligned}$$

where $\sum _{l\in {\mathcal K}} p_l = 1$ holds. The model parameters $\Pi =\bigcup _{j\in {\mathcal {J}}}\Pi _{j}$ and $\{p_l \mid l\in {\mathcal K}\}$ and the posterior probabilities of the true labels $\mu _{il} = \Pr [y_i = l \mid {\mathcal Y}, \Pi ]$ are estimated using the following EM algorithm.

• E-step: for each $i\in {\mathcal I}$, update $\{\mu _{il} \mid l\in {\mathcal K}\}$ as

  $$\begin{aligned} \mu _{il}&= \dfrac{p_l \rho _{il}}{\sum _{l^{\prime }\in {\mathcal K}} p_{l^{\prime }}\rho _{il^{\prime }}},\\ \mathrm{where\ } \log \rho _{il}&= \sum _{j\in {\mathcal {J}}_{i}} \sum _{k\in {\mathcal K}} {\mathbf I}(y_{i,j}=k) \log \pi _{jkl}. \end{aligned}$$

• M-step: for each $j\in {\mathcal {J}}$, update $\Pi _j$ as

  $$\begin{aligned} \pi _{jkl} = \dfrac{\sum _{i\in {\mathcal I}_j} \mu _{il} {\mathbf I}(y_{i,j} = k)}{\sum _{i\in {\mathcal I}_j} \mu _{il}}, \end{aligned}$$

  and for each $l\in {\mathcal K}$, update $p_l$ as

  $$\begin{aligned} p_l = \dfrac{1}{|{\mathcal I}|}\sum _{i\in {\mathcal I}}\mu _{il}. \end{aligned}$$

This algorithm can be extended to preserve worker privacy. In the E-step, the parties calculate $\{\log \rho _{il} \mid i\in {\mathcal I}, l\in {\mathcal K}\}$ using our secure sum protocol, and the requester calculates and broadcasts $\{\mu _{il}\mid i\in {\mathcal I}, l\in {\mathcal K}\}$. In the M-step, each worker $j$ calculates $\{\pi _{jkl} \mid k,l\in {\mathcal K}\}$, and the requester calculates $\{p_l \mid {l\in {\mathcal K}}\}$.

1.2 Appendix 1.2: Real-valued labels

The LC method was modified to deal with real-valued labels by Raykar et al. (2010). Let us assume a task to give a real-valued label. For each $i\in {\mathcal I}$ and $j\in {\mathcal {J}}$, a crowd label $y_{i,j}\in \mathbb {R}$ is generated by the normal distribution

$$\begin{aligned} p(y_{i,j}\mid y_{i}, \tau _j, \gamma ) = \mathcal {N}(y_{i,j} \mid y_{i}, 1/\tau _j + 1/\gamma ), \end{aligned}$$

where $\tau _j (> 0)$ is the precision parameter of the normal distribution, which is interpreted as the ability of worker $j$, and $\gamma$ works as regularization. Let us denote $1/\lambda _j := 1/\tau _j + 1/\gamma$. Assuming that the crowd labels were generated by this model, the true labels and the precision parameters are estimated by the following EM-like algorithm.

• E-step: for each $i\in {\mathcal I}$, update the true label $y_i$ as

  $$\begin{aligned} y_i = \dfrac{\sum _{j\in {\mathcal {J}}_i} \lambda _j y_{i,j}}{\sum _{j\in {\mathcal {J}}_i} \lambda _j}. \end{aligned}$$

• M-step: for each $j\in {\mathcal {J}}$, update $\lambda _j$ by solving

  $$\begin{aligned} \dfrac{1}{\lambda _j} = \dfrac{1}{|{\mathcal I}_j|}\sum _{i\in {\mathcal I}_j} (y_{i,j} - y_i)^2. \end{aligned}$$

This algorithm can also be extended to preserve worker privacy. In the E-step, the parties calculate $\left\{ \sum _{j\in {\mathcal {J}}_i} \lambda _j y_{i,j}, \sum _{j\in {\mathcal {J}}_i} \lambda _j \mid {i\in {\mathcal I}}\right\}$ using our secure sum protocol, and the requester calculates and broadcasts $\{y_i \mid {i\in {\mathcal I}}\}$. In the M-step, each worker $j$ calculates $\lambda _j$.