Search

In the realm of dermatological diagnoses, where the analysis of dermatoscopic and microscopic skin lesion images is pivotal for the accurate and early detection of various medical conditions, the costs associated with creating diverse and high-quality annotated datasets have hampered the accuracy and generalizability of machine learning models. We propose an innovative unsupervised augmentation solution that harnesses Generative Adversarial Network (GAN) based models and associated techniques over their latent space to generate controlled semiautomatically-discovered semantic variations in dermatoscopic images. We created synthetic images to incorporate the semantic variations and augmented the training data with these images. With this approach, we were able to increase the performance of machine learning models and set a new benchmark amongst non-ensemble based models in skin lesion classification on the HAM10000 dataset; and used the observed analytics and generated models for detailed studies on model explainability, affirming the effectiveness of our solution., Comment: This preprint has been submitted to the Workshop on Synthetic Data for Computer Vision (SyntheticData4CV 2024 is a side event on 18th European Conference on Computer Vision 2024). This preprint has not undergone peer review or any post-submission improvements or corrections

Adversarial attacks are small, carefully crafted perturbations, imperceptible to the naked eye; that when added to an image cause deep learning models to misclassify the image with potentially detrimental outcomes. With the rise of artificial intelligence models in consumer safety and security intensive industries such as self-driving cars, camera surveillance and face recognition, there is a growing need for guarding against adversarial attacks. In this paper, we present an approach that uses metamorphic testing principles to automatically detect such adversarial attacks. The approach can detect image manipulations that are so small, that they are impossible to detect by a human through visual inspection. By applying metamorphic relations based on distance ratio preserving affine image transformations which compare the behavior of the original and transformed image; we show that our proposed approach can determine whether or not the input image is adversarial with a high degree of accuracy.

Autonomy is the ability of a system to make decisions and take actions in the presence of uncertainty. Validating autonomy, therefore, is a matter of validating that the system makes intelligent decisions even when there might be discrepancies between the observed and the expected state of the world as well as when the possible outcome of each of the possible courses of action is uncertain. During the requirements engineering phase, the functions, including those that are required to be autonomous and their expected behavior are specified. However, in the case of autonomous software system, it is challenging to specify all possible scenarios that the software may encounter. Additionally, since autonomous behaviors cannot always be fully predetermined, it is difficult to reason about their completeness and correctness. While these challenges are also faced with non-autonomous system, they are more prevalent for autonomous system due to their complexity and emergent behavior. Our efforts have been three-fold. On one hand, we have developed a goal-based method for requirements decomposition. We have embedded this approach into a higher-level systems engineering framework and then developed methods for decision and reliability analysis within this framework. And finally, we use simulation analysis to validate the more theoretical methods and provide a feedback mechanism. The decision analysis approach (DA) helps in identifying and mitigating the run-time risks, by bringing to the forefront the uncertainties, decisions, interactions, and other factors that may cause autonomous software to make erroneous decisions. The DA framework is used to quantify, for each decision that the software can make, failure risk due to uncertainties. The information about the key decisions and circumstances that can cause autonomous system to make incorrect decisions are used to generate test scenarios to be run in a simulator to ensure that the system can handle error-prone circumstances. The results or outcomes from decisions, based on the simulated scenarios, can then be leveraged to further gain insight of the assured system. We combine a goal-based approach with simulation analysis to facilitate requirements development for autonomous systems and further provide a method for run-time systems level reliability considerations during the autonomous software development process. The DA and the reliability analysis methods provide a formalism for the consideration of uncertainty in the run time decision-making process for autonomous software. This formalism considers the uncertainty involved in the outcome of different courses of action, in terms of performance and cost, as well as a method to consider the system level reliability implications. System level reliability considerations for autonomous software decision making contribute to the satisfaction of the system level goals over the lifecycle of a given mission. We demonstrate our approach using a hypothetical rover path planning example. The requirements for our rover simulator are based on existing literature about Mars rovers and lessons learned from real world observations at JPL. The specifics of the autonomy design for the Mars rovers, that have not been previously cleared for external release, are not used for our demonstrations.

We report on a preliminary case study to examine software safety risk in the early design phase of the NASA Constellation spaceflight program. Our goal is to provide NASA quality assurance managers with information regarding the ongoing state of software safety across the program. We examined 154 hazard reports created during the preliminary design phase of three major flight hardware systems within the Constellation program. Our purpose was two-fold: 1) to quantify the relative importance of software with respect to system safety; and 2) to identify potential risks due to incorrect application of the safety process, deficiencies in the safety process, or the lack of a defined process. One early outcome of this work was to show that there are structural deficiencies in collecting valid safety data that make software safety different from hardware safety. In our conclusions we present some of these deficiencies.

A significant body of knowledge concerning software inspection practice indicates that the value of inspections varies widely both within and across organizations. Inspection effectiveness and efficiency can be measured in numerous ways, and may be affected by a variety of factors such as Inspection planning, the type of software, the developing organization, and many others. In the early 1990's, NASA formulated heuristics for inspection planning based on best practices and early NASA inspection data. Over the intervening years, the body of data from NASA inspections has grown. This paper describes a multi-faceted exploratory analysis performed on this · data to elicit lessons learned in general about conducting inspections and to recommend improvements to the existing heuristics. The contributions of our results include support for modifying some of the original inspection heuristics (e.g. Increasing the recommended page rate), evidence that Inspection planners must choose between efficiency and effectiveness, as a good tradeoff between them may not exist, and Identification of small subsets of inspections for which new inspection heuristics are needed. Most Importantly, this work illustrates the value of collecting rich data on software Inspections, and using it to gain insight into, and Improve, inspection practice.