Abstract
Phishing attacks aims to steal sensitive information and, unfortunately, are becoming a common practice on the web. Email phishing is one of the most common types of attacks on the web and can have a big impact on individuals and enterprises. There is still a gap in prevention when it comes to detecting phishing emails, as new attacks are usually not detected. The goal of this work was to develop a model capable of identifying phishing emails based on machine learning approaches. The work was performed in collaboration with E-goi, a multi-channel marketing automation company. The data consisted of emails collected from the E-goi servers in the electronic mail format. The problem consisted of a classification problem with unbalanced classes, with the minority class corresponding to the phishing emails and having less than 1% of the total emails. Several models were evaluated after careful data selection and feature extraction based on the email content and the literature regarding these types of problems. Due to the imbalance present in the data, several sampling methods based on under-sampling techniques were tested to see their impact on the model's ability to detect phishing emails. The final model consisted of a neural network able to detect more than 80% of phishing emails without compromising the remaining emails sent by E-goi clients.

Abstract
Object detection is a crucial task in autonomous driving, where domain shift between the training and the test set is one of the main reasons behind the poor performance of a detector when deployed. Some erroneous priors may be learned from the training set, therefore a model must be invariant to conditions that might promote such priors. To tackle this problem, we propose an adversarial learning framework consisting of an encoder, an object-detector, and a condition-classifier. The encoder is trained to deceive the condition-classifier and aid the object-detector as much as possible throughout the learning stage, in order to obtain highly discriminative features. Experiments showed that this framework is not very competitive regarding the trade-off between precision and recall, but it does improve the ability of the model to detect smaller objects and some object classes.

Abstract
Autonomous driving can reduce the number of road accidents due to human error and result in safer roads. One important part of the system is the perception unit, which provides information about the environment surrounding the car. Currently, most manufacturers are using not only RGB cameras, which are passive sensors that capture light already in the environment but also Lidar. This sensor actively emits laser pulses to a surface or object and measures reflection and time-of-flight. Previous work, YOLOP, already proposed a model for object detection and semantic segmentation, but only using RGB. This work extends it for Lidar and evaluates performance on KITTI, a public autonomous driving dataset. The implementation shows improved precision across all objects of different sizes. The implementation is entirely made available: https://github.com/filipepcampos/yolomm.

Abstract
After the learning process, certain types of images may not be modeled correctly because they were not well represented in the training set. These failures can then be compensated for by collecting more images from the real-world and incorporating them into the learning process - an expensive process known as active learning. The proposed twist, called active supervision, uses the model itself to change the existing images in the direction where the boundary is less defined and requests feedback from the user on how the new image should be labeled. Experiments in the context of class imbalance show the technique is able to increase model performance in rare classes. Active human supervision helps provide crucial information to the model during training that the training set lacks.

Abstract
This paper introduces an evaluation procedure to validate the efficacy of explanation methods for Convolutional Neural Network (CNN) models in ordinal regression tasks. Two ordinal methods are contrasted against a baseline using cross-entropy, across four datasets. A statistical analysis demonstrates that attribution methods, such as Grad-CAM and IBA, perform significantly better when used with ordinal regression CNN models compared to a baseline approach in most ordinal and nominal metrics. The study suggests that incorporating ordinal information into the attribution map construction process may improve the explanations further.

Abstract
Semantic segmentation consists of classifying each pixel according to a set of classes. This process is particularly slow for high-resolution images, which are present in many applications, ranging from biomedicine to the automotive industry. In this work, we propose an algorithm targeted to segment high-resolution images based on two stages. During stage 1, a lower-resolution interpolation of the image is the input of a first neural network, whose low-resolution output is resized to the original resolution. Then, in stage 2, the probabilities resulting from stage 1 are divided into contiguous patches, with less confident ones being collected and refined by a second neural network. The main novelty of this algorithm is the aggregation of the low-resolution result from stage 1 with the high-resolution patches from stage 2. We propose the U-Net architecture segmentation, evaluated in six databases. Our method shows similar results to the baseline regarding the Dice coefficient, with fewer arithmetic operations.