Data leakage is a very common problem that is often overlooked during splitting data into train and test sets before training any ML/DL model. The model performance gets artificially inflated with the presence of data leakage during the evaluation phase which often leads the model to erroneous prediction on real-time deployment. However, detecting the presence of such leakage is challenging, particularly in the object detection context of perception systems where the model needs to be supplied with image data for training. In this study, we conduct a computational experiment to develop a method for detecting data leakage. We then conducted an initial evaluation of the method as a first step on a public dataset, “Kitti”, which is a popular and widely accepted benchmark dataset in the automotive domain. The evaluation results show that our proposed D-LeDe method are able to successfully detect potential data leakage caused by image similarity. A further validation was also provided to justify the evaluation outcome by conducting pair-wise image similarity analysis using perceptual hash (pHash) distance.

Deep learning (DL) systems are becoming an essential part of software systems, so it is necessary to test them thoroughly. This is a challenging task since the test sets can grow over time as the new data is being acquired, and it becomes time-consuming. Input prioritization is necessary to reduce the testing time since prioritized test inputs are more likely to reveal the erroneous behavior of a DL system earlier during test execution. Input prioritization approaches have been rudimentary analyzed against each other, this study compares different input prioritization techniques regarding their effectiveness and efficiency. This work considers surprise adequacy, autoencoder-based, and similarity-based input prioritization approaches in the example of testing a DL image classification algorithms applied on MNIST, Fashion-MNIST, CIFAR-10, and STL-10 datasets. To measure effectiveness and efficiency, we use a modified APFD (Average Percentage of Fault Detected), and set up & execution time, respectively. We observe that the surprise adequacy is the most effective (0.785 to 0.914 APFD). The autoencoder-based and similarity-based techniques are less effective, with the performance from 0.532 to 0.744 APFD and 0.579 to 0.709 APFD, respectively. In contrast, the similarity-based and surprise adequacy-based approaches are the most and least efficient, respectively. The findings in this work demonstrate the trade-off between the considered input prioritization techniques to understanding their practical applicability for testing DL algorithms.

In recent years, with more edge devices being put into use, the amount of data that is created, transmitted and stored is increasing exponentially. Moreover, due to the development of machine learning algorithms, modern software-intensive systems are able to take advantage of the data to further improve their service quality. However, it is expensive and inefficient to transmit large amounts of data to a central location for the purpose of training and deploying machine learning models. Data transfer from edge devices across the globe to central locations may also raise privacy and concerns related to local data regulations. As a distributed learning approach, Federated Learning has been introduced to tackle those challenges. Since Federated Learning simply exchanges locally trained machine learning models rather than the entire data set throughout the training process, the method not only protects user data privacy but also improves model training efficiency. In this paper, we have investigated an advanced machine learning algorithm, Deep Neural Decision Forests (DNDF), which unites classification trees with the representation learning functionality from deep convolutional neural networks. In this paper, we propose a novel algorithm, AF-DNDF which extends DNDF with an asynchronous federated aggregation protocol. Based on the local quality of each classification tree, our architecture can select and combine the optimal groups of decision trees from multiple local devices. The introduction of the asynchronous protocol enables the algorithm to be deployed in the industrial context with heterogeneous hardware settings. Our AF-DNDF architecture is validated in an automotive industrial use case focusing on road objects recognition and demonstrated by an empirical experiment with two different data sets. The experimental results show that our AF-DNDF algorithm significantly reduces the communication overhead and accelerates model training speed without sacrificing model classification performance. The algorithm can reach the same classification accuracy as the commonly used centralized machine learning methods but also greatly improve local edge model quality.

Applications for adaptive (sometimes also called smart) Cyber-Physical Systems are blossoming thanks to the large volumes of data, sensed in a continuous fashion, in large distributed systems. The benefits of these applications come nonetheless with a price: the need for jointly addressing challenges in efficient data communication and analysis (among others). The goal of the DRIVEN framework, presented here, is to address these challenges for a data gathering and distance-based clustering tool in the context of vehicular networks. Because of the limited communication bandwidth (compared to the volume of sensed data) of vehicular networks and the monetary costs of data transmission, the intuition behind DRIVEN is to avoid gathering the data to be clustered in a raw format from each vehicle, but rather to allow for a streaming-based error-bounded approximation, through Piecewise Linear Approximation, to compress the volumes of data to be gathered. At the same time, rather than relying on a batch-based clustering algorithm that requires all the data to be first gathered (and then clustered), DRIVEN relies on and extends a streaming-based clustering algorithm that leverages the inherent ordering of the spatial and temporal data being collected, to perform the clustering in an online fashion, while data is being retrieved. As we show, based on our prototype implementation using Apache Flink and our evaluation with real-world data such as GPS and LiDAR, the accuracy loss for the clustering performed on the reconstructed data can be small, even when the raw data is compressed to 10-35% of its original size, and the transferring of data itself can be completed in up to one-tenth of the duration observed when gathering raw data.