Browsing by Author "Pérez, Aritz"

Now showing 1 - 3 of 3
  • No Thumbnail Available
    Item
    On-Line Dynamic Time Warping for Streaming Time Series
    (Springer Verlag, 2017) Oregi, Izaskun; Pérez, Aritz; Del Ser, Javier; Lozano, José A.; Ceci, Michelangelo; Hollmen, Jaakko; Todorovski, Ljupco; Vens, Celine; Dzeroski, Saso; Quantum; IA
    Dynamic Time Warping is a well-known measure of dissimilarity between time series. Due to its flexibility to deal with non-linear distortions along the time axis, this measure has been widely utilized in machine learning models for this particular kind of data. Nowadays, the proliferation of streaming data sources has ignited the interest and attention of the scientific community around on-line learning models. In this work, we naturally adapt Dynamic Time Warping to the on-line learning setting. Specifically, we propose a novel on-line measure of dissimilarity for streaming time series which combines a warp constraint and a weighted memory mechanism to simplify the time series alignment and adapt to non-stationary data intervals along time. Computer simulations are analyzed and discussed so as to shed light on the performance and complexity of the proposed measure.
  • No Thumbnail Available
    Item
    On-line Elastic Similarity Measures for time series
    (2019-04) Oregi, Izaskun; Pérez, Aritz; Del Ser, Javier; Lozano, Jose A.; Quantum; IA
    The way similarity is measured among time series is of paramount importance in many data mining and machine learning tasks. For instance, Elastic Similarity Measures are widely used to determine whether two time series are similar to each other. Indeed, in off-line time series mining, these measures have been shown to be very effective due to their ability to handle time distortions and mitigate their effect on the resulting distance. In the on-line setting, where available data increase continuously over time and not necessary in a stationary manner, stream mining approaches are required to be fast with limited memory consumption and capable of adapting to different stationary intervals. In this sense, the computational complexity of Elastic Similarity Measures and their lack of flexibility to accommodate different stationary intervals, make these similarity measures incompatible with the requirements mentioned. To overcome these issues, this paper adapts the family of Elastic Similarity Measures – which includes Dynamic Time Warping, Edit Distance, Edit Distance for Real Sequences and Edit Distance with Real Penalty – to the on-line setting. The proposed adaptation is based on two main ideas: a forgetting mechanism and the incremental computation. The former makes the similarity consistent with streaming time series characteristics by giving more importance to recent observations, whereas the latter reduces the computational complexity by avoiding unnecessary computations. In order to assess the behavior of the proposed similarity measure in on-line settings, two different experiments have been carried out. The first aims at showing the efficiency of the proposed adaptation, to do so we calculate and compare the computation time for the elastic measures and their on-line adaptation. By analyzing the results drawn from a distance-based streaming machine learning model, the second experiment intends to show the effect of the forgetting mechanism on the resulting similarity value. The experimentation shows, for the aforementioned Elastic Similarity Measures, that the proposed adaptation meets the memory, computational complexity and flexibility constraints imposed by streaming data.
  • No Thumbnail Available
    Item
    Robust image classification against adversarial attacks using elastic similarity measures between edge count sequences
    (2020-08) Oregi, Izaskun; Del Ser, Javier; Pérez, Aritz; Lozano, José A.; Quantum; IA
    Due to their unprecedented capacity to learn patterns from raw data, deep neural networks have become the de facto modeling choice to address complex machine learning tasks. However, recent works have emphasized the vulnerability of deep neural networks when being fed with intelligently manipulated adversarial data instances tailored to confuse the model. In order to overcome this issue, a major effort has been made to find methods capable of making deep learning models robust against adversarial inputs. This work presents a new perspective for improving the robustness of deep neural networks in image classification. In computer vision scenarios, adversarial images are crafted by manipulating legitimate inputs so that the target classifier is eventually fooled, but the manipulation is not visually distinguishable by an external observer. The reason for the imperceptibility of the attack is that the human visual system fails to detect minor variations in color space, but excels at detecting anomalies in geometric shapes. We capitalize on this fact by extracting color gradient features from input images at multiple sensitivity levels to detect possible manipulations. We resort to a deep neural classifier to predict the category of unseen images, whereas a discrimination model analyzes the extracted color gradient features with time series techniques to determine the legitimacy of input images. The performance of our method is assessed over experiments comprising state-of-the-art techniques for crafting adversarial attacks. Results corroborate the increased robustness of the classifier when using our discrimination module, yielding drastically reduced success rates of adversarial attacks that operate on the whole image rather than on localized regions or around the existing shapes of the image. Future research is outlined towards improving the detection accuracy of the proposed method for more general attack strategies.