130 lines
11 KiB
TeX
130 lines
11 KiB
TeX
\section{Evaluation}
|
|
\label{sec:lmdk-eval}
|
|
|
|
\kat{After discussing with Dimitris, I thought you are keeping one chapter for the proposals of the thesis. In this case, it would be more clean to keep the theoretical contributions in one chapter and the evaluation in a separate chapter. }
|
|
In this section we present the experiments that we performed on real and synthetic data sets.
|
|
With the experiments on the synthetic data sets we show the privacy loss by our framework when tuning the size and statistical characteristics of the input {\thething} set $L$.
|
|
We also show how the privacy loss under temporal correlation is affected by the number and distribution of the {\thethings}.
|
|
With the experiments on the real data sets, we show the performance in terms of utility of our three {\thething} mechanisms.
|
|
|
|
Notice that in our experiments, in the cases when we have $0\%$ and $100\%$ of the events being {\thethings}, we get the same behavior as in event- and user-level privacy respectively.
|
|
This happens due the fact that at each timestamp we take into account only the data items at the current timestamp and ignore the rest of the time series (event-level) when there are no {\thethings}.
|
|
Whereas, when each timestamp corresponds to a {\thething} we consider and protect all the events throughout the entire series (user-level).
|
|
|
|
|
|
\subsection{Setting, configurations, and data sets}
|
|
\paragraph{Setting}
|
|
We implemented our experiments\footnote{Code available at \url{https://gitlab.com/adhesivegoldfinch/cikm}} in Python $3$.$9$.$5$ and executed them on a machine with Intel i$7$-$6700$HQ $3$.$5$GHz CPU and $16$GB RAM, running Manjaro $21$.$0$.$5$.
|
|
We repeated each experiment $100$ times and we report the mean over these iterations.
|
|
|
|
|
|
\paragraph{Data sets}
|
|
For the \emph{real} data sets, we used the Geolife~\cite{zheng2010geolife} and T-drive~\cite{yuan2010t} from which we sampled the first $1000$ data items.
|
|
We achieved the desired {\thethings} percentages by utilizing the method of Li et al.~\cite{li2008mining} for detecting stay points in trajectory data.
|
|
In more detail, the algorithm checks for each data item if each subsequent item is within a given distance threshold $\Delta l$ and measures the time period $\Delta t$ between the present point and the last subsequent point.
|
|
We achieve $0$, $20$ $40$, $60$, $80$, and $100$ {\thethings} percentages by setting the ($\Delta l$ in meters, $\Delta t$ in minutes) pairs input to the stay point discovery method for T-drive as [($0$, $1000$), ($2095$, $30$), ($2790$, $30$), ($3590$, $30$), ($4825$, $30$), ($10350$, $30$)] and for Geolife as [($0$, $100000$), ($205$, $30$), ($450$, $30$), ($725$, $30$), ($855$, $30$), ($50000$, $30$)].
|
|
|
|
|
|
Next, we generated synthetic time series of length equal to $100$ timestamps, for which we varied the number and distribution of {\thethings}.
|
|
% to achieve the necessary {\thethings} distribution and percentage for where applicable.
|
|
% \paragraph{{\Thethings} distribution}
|
|
We created \emph{left-skewed} (the {\thethings} are distributed towards the end), \emph{symmetric} (in the middle), \emph{right-skewed} (in the beginning), \emph{bimodal} (both end and beginning), and \emph{uniform} (all over the time series) {\thething} distributions.
|
|
%, in the beginning and in the end (\emph{bimodal}), and all over the extend (\emph{uniform}) of a time series.
|
|
When pertinent, we group the left- and right-skewed cases as simply `skewed', since they share several features due to symmetry.
|
|
In order to get {\thethings} with the above distribution features, we generate probability distributions with appropriate characteristics and sample from them, without replacement, the desired number of points.
|
|
%The generated distributions are representative of the cases that we wish to examine during the experiments.
|
|
% For example, for a left-skewed {\thethings} distribution we would utilize a truncated distribution resulting from the restriction of the domain of a normal distribution to the beginning and end of the time series with its location shifted to the center of the right half of the series.
|
|
For consistency, we calculate the scale parameter depending on the length of the series by setting it equal to the series' length over a constant.
|
|
%We take into account only the temporal order of the points and the position of regular and {\thething} events within the series.
|
|
Note, that for the experiments performed on the synthetic data sets, the original values to be released do not influence the outcome of our conclusions, thus we ignore them.
|
|
|
|
|
|
\paragraph{Configurations}
|
|
We model the temporal correlation in the synthetic data as a \emph{stochastic matrix} $P$, using a \emph{Markov Chain}~\cite{gagniuc2017markov}.
|
|
$P$ is a $n \times n$ matrix, where the element $p_{ij}$
|
|
%at the $i$th row of the $j$th column that
|
|
represents the transition probability from a state $i$ to another state $j$.
|
|
%, $\forall i, j \leq n$.
|
|
It holds that the elements of every row $j$ of $P$ sum up to $1$.
|
|
We follow the \emph{Laplacian smoothing} technique~\cite{sorkine2004laplacian} as utilized in~\cite{cao2018quantifying} to generate the matrix $P$ with a degree of temporal correlation $s>0$.
|
|
% and generate a stochastic matrix $P$ with a degree of temporal correlation $s$ by calculating each element $P_{ij}$ as follows
|
|
%$$\frac{(I_{n})_{ij} + s}{\sum_{k = 1}^{n}((I_{n})_{jk} + s)}$$
|
|
%where $I_{n}$ is an \emph{identity matrix} of size $n$.
|
|
%, i.e.,~an $n \times n$ matrix with $1$s on its main diagonal and $0$s elsewhere.
|
|
% $s$ takes only positive values which are comparable only for stochastic matrices of the same size.
|
|
$s$ dictates the strength of the correlation; the lower its value,
|
|
%the lower the degree of uniformity of each row, and therefore
|
|
the stronger the correlation degree.
|
|
%In general, larger transition matrices tend to be uniform, resulting in weaker correlation.
|
|
In our experiments, for simplicity, we set $n = 2$ and we investigate the effect of \emph{weak} ($s = 1$), \emph{moderate} ($s = 0.1$), and \emph{strong} ($s = 0.01$) temporal correlation degree on the overall privacy loss.
|
|
|
|
We set $\varepsilon = 1$.
|
|
To perturb the spatial values of the real data sets, we inject noise that we sample from the Planar Laplace mechanism~\cite{andres2013geo}.
|
|
Finally, notice that all diagrams are in logarithmic scale.
|
|
|
|
\subsection{Experiments}
|
|
|
|
\paragraph{Budget allocation schemes}
|
|
|
|
Figure~\ref{fig:real} exhibits the performance of the three mechanisms: Skip, Uniform, and Adaptive.
|
|
|
|
\begin{figure}[htp]
|
|
\centering
|
|
\subcaptionbox{Geolife\label{fig:geolife}}{%
|
|
\includegraphics[width=.5\linewidth]{geolife}%
|
|
}%
|
|
\subcaptionbox{T-drive\label{fig:t-drive}}{%
|
|
\includegraphics[width=.5\linewidth]{t-drive}%
|
|
}%
|
|
\caption{The mean absolute error (in meters) of the released data for different {\thethings} percentages.}
|
|
\label{fig:real}
|
|
\end{figure}
|
|
|
|
For the Geolife data set (Figure~\ref{fig:geolife}), Skip has the best performance (measured in Mean Absolute Error, in meters) because it invests the most budget overall at every regular event, by approximating the {\thething} data based on previous releases.
|
|
Due to the data set's high density (every $1$--$5$ seconds or every $5$--$10$ meters per point) approximating constantly has a low impact on the data utility.
|
|
On the contrary, the lower density of the T-drive data set (Figure~\ref{fig:t-drive}) has a negative impact on the performance of Skip.
|
|
In the T-drive data set, the Adaptive mechanism outperforms the Uniform one by $10$\%--$20$\% for all {\thethings} percentages greater than $0$ and by more than $20$\% the Skip one.
|
|
In general, we can claim that the Adaptive is the best performing mechanism, if we take into consideration the drawbacks of the Skip mechanism mentioned in Section~\ref{subsec:lmdk-mechs}. Moreover, designing a data-dependent sampling scheme would possibly result in better results for Adaptive.
|
|
|
|
|
|
\paragraph{Temporal distance and correlation}
|
|
Figure~\ref{fig:avg-dist} shows a comparison of the average temporal distance of the events from the previous/next {\thething} or the start/end of the time series for various distributions in synthetic data.
|
|
More particularly, we count for every event the total number of events between itself and the nearest {\thething} or the series edge.
|
|
We observe that the uniform and bimodal distributions tend to limit the regular event--{\thething} distance.
|
|
This is due to the fact that the former scatters the {\thethings}, while the latter distributes them on both edges, leaving a shorter space uninterrupted by {\thethings}.
|
|
% and as a result they reduce the uninterrupted space by landmarks in the sequence.
|
|
On the contrary, distributing the {\thethings} at one part of the sequence, as in skewed or symmetric, creates a wider space without {\thethings}.
|
|
|
|
\begin{figure}[htp]
|
|
\centering
|
|
\includegraphics[width=.5\linewidth]{avg-dist}%
|
|
\caption{Average temporal distance of the events from the {\thethings} for different {\thethings} percentages within a time series in various {\thethings} distributions.}
|
|
\label{fig:avg-dist}
|
|
\end{figure}
|
|
|
|
Figure~\ref{fig:dist-cor} illustrates a comparison among the aforementioned distributions regarding the overall privacy loss under moderate (Figure~\ref{fig:dist-cor-mod}), and strong (Figure~\ref{fig:dist-cor-stg}) correlation degrees.
|
|
The line shows the overall privacy loss---for all cases of {\thethings} distribution---without temporal correlation.
|
|
We skip the presentation of the results under a weak correlation degree, since they converge in this case.
|
|
In combination with Figure~\ref{fig:avg-dist}, we conclude that a greater average event-{\thething} distance in a distribution can result into greater overall privacy loss under moderate and strong temporal correlation.
|
|
This is due to the fact that the backward/forward privacy loss accumulates more over time in wider spaces without {\thethings} (see Section~\ref{subsec:correlations}).
|
|
Furthermore, the behavior of the privacy loss is as expected regarding the temporal correlation degree.
|
|
Predictably, a stronger correlation degree generates higher privacy loss while widening the gap between the different distribution cases.
|
|
On the contrary, a weaker correlation degree makes it harder to differentiate among the {\thethings} distributions.
|
|
|
|
\begin{figure}[htp]
|
|
\centering
|
|
\subcaptionbox{Weak correlation\label{fig:dist-cor-wk}}{%
|
|
\includegraphics[width=.5\linewidth]{dist-cor-wk}%
|
|
}%
|
|
\hspace{\fill}
|
|
\subcaptionbox{Moderate correlation\label{fig:dist-cor-mod}}{%
|
|
\includegraphics[width=.5\linewidth]{dist-cor-mod}%
|
|
}%
|
|
\subcaptionbox{Strong correlation\label{fig:dist-cor-stg}}{%
|
|
\includegraphics[width=.5\linewidth]{dist-cor-stg}%
|
|
}%
|
|
\caption{Privacy loss for different {\thethings} percentages and distributions, under weak, moderate, and strong degrees of temporal correlation.
|
|
The line shows the overall privacy loss without temporal correlation.}
|
|
\label{fig:dist-cor}
|
|
\end{figure}
|