|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Multimedia content description and structuring
The vulnerability of learning to adversarial perturbation increases with intrinsic dimensionality
Participant : Laurent Amsaleg.
Joint work with James Bailey, Dominique Barbe, Sarah Erfani, Michael Houle, Vinh Nguyen amd Miloš Radovanovic.
Recent research has shown that machine learning systems, including state-of-the-art deep neural networks, are vulnerable to adversarial attacks. By adding to the input object an imperceptible amount of adversarial noise, it is highly likely that the classifier can be tricked into assigning the modified object to any desired class. It has also been observed that these adversarial samples generalize well across models. A complete understanding of the nature of adversarial samples has not yet emerged. Towards this goal, we present a novel theoretical result formally linking the adversarial vulnerability of learning to the intrinsic dimensionality of the data. In particular, our investigation establishes that as the local intrinsic dimensionality (LID) increases, 1-NN classifiers become increasingly prone to being subverted. We show that in expectation, a k-nearest neighbor of a test point can be transformed into its 1-nearest neighbor by adding an amount of noise that diminishes as the LID increases. We also provide an experimental validation of the impact of LID on adversarial perturbation for both synthetic and real data, and discuss the implications of our result for general classifiers [13].
Efficient temporal kernels between feature sets for time series classification
Participant : Simon Malinowski.
Joint work with Romain Tavenard, Adeline Bailly, Louis Chapel (Univ. Rennes 2), Benjamin Bustos and Heider Sanchez (Univ. of Chile).
In the time-series classification context, the majority of the most accurate core methods are based on the bag-of-words framework, in which sets of local features are first extracted from time series. A dictionary of words is then learned and each time series is finally represented by a histogram of word occurrences. This representation induces a loss of information due to the quantization of features into words as all the time series are represented using the same fixed dictionary. In order to overcome this issue, we have designed a kernel operating directly on sets of features. Then, we have extended it to a time-compliant kernel that allows one to take into account the temporal information. We applied this kernel in the time series classification context. Proposed kernel has a quadratic complexity with the size of input feature sets, which is problematic when dealing with long time series. However, we have shown that kernel approximation techniques can be used to define a good trade-off between accuracy and complexity. We experimentally demonstrated that the proposed kernel can significantly improve the performance of time series classification algorithms based on bag-of-words [33].
Tampering detection and localization in images from social networks
Participants : Cédric Maigrot, Ewa Kijak, Vincent Claveau.
Verifying the authenticity of an image broadcast on social networks is crucial to limit the dissemination of false information. In this work, we aim to provide information about tampering localisation on such images, in order to help either the user or automatic methods to discriminate truth from falsehood. These images may have been subjected to a large number of possible forgeries, which calls for the use of generic methods. Image forensics methods based on local features have proven to be effective for the specific case of copy-move forgery. By taking advantage of the number of images available on the internet, we propose a generic system based on image retrieval, and image comparison based on local features to localise any kind of tampering in images from social networks.
Images from social media are likely to have undergone a large variety of modifications, some being malicious, and some not. The proposed approach is evaluated on three dedicated datasets containing a variety of representative tamperings in images from social media, with difficult examples. This allows an analysis of the local-features approaches behavior in this context. The method is further compared to several state-of-the-art methods and proves to be superior. Finally, we propose a classification step to discriminate malicious modifications from the non-malicious ones.
We have also built and made publicly available a large and challenging adapted database of real case images for evaluation [29].
Identity documents classification as an image classification problem
Participants : Ronan Sicre, Teddy Furon.
Joint work with Ahmad Montaser Awal and Nabil Ghanni (AriadNext).
This works studies the classification of images of identification documents. More specifically, we address the classification of documents composed of few textual information and complex background (such as identity documents). Unlike most existing systems, the proposed approach simultaneously locates the document and recognizes its class. The latter is defined by the document nature (passport, ID, etc.), emission country, version, and the visible side (main or back). This task is very challenging due to unconstrained capturing conditions, sparse textual information, and varying components that are irrelevant to the classification, e.g. photo, names, address, etc. First, a base of document models is created from reference images.
This problem is critical in various security context where proposed system must offer high performances. We address this challenge as an image classification problem, which has received a large attention from the scientific community. We show that training images are not necessary and only one reference image is enough to create a document model. Then, the query image is matched against all models in the base. Unknown documents are rejected using an estimated quality based on the extracted document. The matching process is optimized to guarantee an execution time independent from the number of document models. Once the document model is found, a more accurate matching is performed to locate the document and facilitate information extraction. Our system is evaluated on several datasets with up to 3042 real documents (representing 64 classes) achieving an accuracy of 96.6 % in [14].
In a second step, several methods are evaluated and we report results allowing a better understanding of the specificity of identification documents. We are especially interested in deep learning approaches, showing good transfer capabilities and high performances [44], [49].
Sentiment analysis
Participants : Vincent Claveau, Christian Raymond.
In the framework of the NexGenTV project, we have participated to the text-mining challenge DeFT about sentiment analysis. We have proposed methods for the identification of figurative language (irony, humor...), and for the classification of figurative and non-figurative tweets according to their polarity. For these tasks, we explore the use of three methods of increasing complexity: i) k-nearest neighbors with information retrieval based techniques, ii) boosting of decision trees, iii) recurrent neural networks [36]. It allows us to evaluate the precise interest of each of our approach and the data representation that they use: bag-of-words for the first one, n-grams for the second and word embedding for the latest.