Differential privacy allows bounding the influence that training data records have on a machine learning model. To use differential privacy in machine learning, data scientists must choose privacy parameters (ϵ, δ ). Choosing meaningful privacy parameters is key, since models trained with weak privacy parameters might result in excessive privacy leakage, while strong privacy parameters might overly degrade model utility. However, privacy parameter values are difficult to choose for two main reasons. First, the theoretical upper bound on privacy loss (ϵ, δ) might be loose, depending on the chosen sensitivity and data distribution of practical datasets. Second, legal requirements and societal norms for anonymization often refer to individual identifiability, to which (ϵ, δ ) are only indirectly related. We transform (ϵ, δ ) to a bound on the Bayesian posterior belief of the adversary assumed by differential privacy concerning the presence of any record in the training dataset. The bound holds for multidimensional queries under composition, and we show that it can be tight in practice. Furthermore, we derive an identifiability bound, which relates the adversary assumed in differential privacy to previous work on membership inference adversaries. We formulate an implementation of this differential privacy adversary that allows data scientists to audit model training and compute empirical identifiability scores and empirical (ϵ, δ ).

The availability of individual load profiles per household in the smart grid end-user domain combined with non-intrusive load monitoring to infer personal data from these load curves has led to privacy concerns. Privacy-enhancing technologies have been proposed to address these concerns. In this paper, the extension of privacy-enhancing technologies by wavelet-based multi-resolution analysis (MRA) is proposed to enhance the options available on the user side. For three types of privacy methods (secure aggregation, masking and differential privacy), we show that MRA not only enhances privacy, but also adds additional flexibility and control for the end-user. The combination of MRA and PETs is evaluated in terms of privacy, computational demands, and real-world feasibility for each of the three method types.

In liberalized energy markets, matching consumption patterns to energy tariffs is desirable, but practically limited due to privacy concerns, both on the side of the consumer and on the side of the utilities. We propose a protocol through which a customer can obtain a better tariff with the help of their smart meter and a third party, based on privacy-preserving load profile matching. Our security analysis shows that the protocol preserves consumer privacy, i.e., neither the load profile nor the matching result are disclosed to the utility, unless the consumer later decides to actually purchase the tariff. In addition, the utility’s load profiles used for matching remain private, allowing each utility to offer special tariffs without disclosing the associated load profiles to their competitors. Our approach is shown to have a smaller ciphertext size than homomorphic encryption in practically relevant configurations. However, matching is only possible with up to about 98 % accuracy in general and 93.5 % based on real-world load profiles, respectively. Depending on the practical requirements, two protocol parameters provide a tradeoff between matching accuracy and ciphertext size.

Decreasing time resolution is the simplest possible privacy enhancing technique for energy consumption data. However, its impact on privacy analyses of load signals has never been studied systematically. Non-intrusive appliance load monitoring algorithms (NIALM) have originally been designed for energy disaggregation for subsequent energy feedback. However, the information on appliance use may also be misused for the extraction of personal information. In this work, the effect of decreasing the time resolution in the usual first step, namely edge detection, is studied. It is shown that event values can be estimated rather reliably, but the detection rate of events significantly decreases with increasing measurement time interval.

Privacy has become a critical topic in the engineering of electric systems. This work proposes an approach for smart-grid-specific privacy requirements engineering by extending previous general privacy requirements engineering frameworks. The proposed extension goes one step further by focusing on privacy in the smart grid. An alignment of smart grid privacy requirements, dependability issues and privacy requirements engineering methods is presented. Starting from this alignment a Threat Tree Analysis is performed to obtain a first set of generic, high level privacy requirements. This set is formulated mostly on the data instead of the information level and provides the basis for further project-specific refinement.

Video footage of real crowded scenes still poses severe challenges for automated surveillance. This paper evaluates clustering methods for finding independent dominant motion fields for an observation period based on a recently published real-time optical flow algorithm. We focus on self-tuning spectral clustering and Isomap combined with k-means. Several combinations of feature vector normalizations and distance measures (Euclidean, Mahanalobis and a general additive distance) are evaluated for four image sequences including three publicly available crowd datasets. Evaluation is based on mean accuracy obtained by comparison with a manually defined ground truth clustering. For every dataset at least one approach correctly classified more than 95% of the flow vectors without extra tuning of parameters, providing a basis for an automatic analysis after a view-dependent setup.

AdaBoost.M2 is a boosting algorithm designed for multiclass problems with weak base classifiers. The algorithm is designed to minimize a very loose bound on the training error. We propose two alternative boosting algorithms which also minimize bounds on performance measures. These performance measures are not as strongly connected to the expected error as the training error, but the derived bounds are tighter than the bound on the training error of AdaBoost.M2. In experiments the methods have roughly the same performance in minimizing the training and test error rates. The new algorithms have the advantage that the base classifier should minimize the confidence-rated error, whereas for AdaBoost.M2 the base classifier should minimize the pseudo-loss. This makes them more easily applicable to already existing base classifiers. The new algorithms also tend to converge faster than AdaBoost.M2.